EU AI Act Audit Evidence Pack
AI Agent Detection Rule Set — Conformity Assessment Document
Audit Signature
| Pack SHA-256 | ef437f8f184a834829bc9e3be565662d3849f29d963fae302922f8f86748d7ed |
| Rules Merkle Root | f1c558fdd9c38230be48a3c5855100862dc99b5d0c9ff285048388c2df29a6ad |
| Algorithm | sha256 |
| Signed at | 2026-05-04T15:02:42.037Z |
Auditor verification: recompute the pack SHA-256 over the
JSON form of this pack with the signature block excluded; recompute
the merkle root by SHA-256-pairing the per-rule hashes (sorted by
rule_id). Both should match the values above. The companion JSON
file (eu-ai-act-evidence-pack.json) is the canonical input.
Coverage Summary
| Total rules in scope | 50 |
| Rules with AI agent analogue (production-grade) | 40 |
| Rules needing human review | 10 |
| Hand-crafted (LLM-enriched) rules | 40 |
| Placeholder rules (auto-generated, pending review) | 10 |
Framework Mapping Coverage
EU AI Act
Rules with at least one eu_ai_act mapping: 50
| Reference | Rules | Primary | Secondary | Partial |
| Article 12 | 18 | 7 | 1 | 10 |
| Article 14 | 13 | 13 | 0 | 0 |
| Article 15 | 40 | 40 | 0 | 0 |
| Article 50 | 4 | 2 | 2 | 0 |
| Article 9 | 4 | 4 | 0 | 0 |
OWASP Agentic Top 10 (2026)
Rules with at least one owasp_agentic mapping: 40
| Reference | Rules | Primary | Secondary | Partial |
| ASI01:2026 | 3 | 1 | 2 | 0 |
| ASI02:2026 | 4 | 4 | 0 | 0 |
| ASI04:2026 | 2 | 2 | 0 | 0 |
| ASI05:2026 | 17 | 17 | 0 | 0 |
| ASI06:2026 | 21 | 18 | 3 | 0 |
| ASI07:2026 | 4 | 3 | 1 | 0 |
| ASI08:2026 | 1 | 1 | 0 | 0 |
| ASI09:2026 | 3 | 2 | 1 | 0 |
OWASP LLM Top 10 (2025)
Rules with at least one owasp_llm mapping: 36
| Reference | Rules | Primary | Secondary | Partial |
| LLM01:2025 | 4 | 3 | 1 | 0 |
| LLM02:2025 | 11 | 11 | 0 | 0 |
| LLM03:2025 | 3 | 3 | 0 | 0 |
| LLM06:2025 | 19 | 15 | 3 | 1 |
NIST AI RMF 1.0
Rules with at least one nist_ai_rmf mapping: 50
| Reference | Rules | Primary | Secondary | Partial |
| Govern / GV.1.1 | 10 | 0 | 0 | 10 |
| Manage / MG.2.3 | 18 | 18 | 0 | 0 |
| Manage / MG.3.2 | 2 | 2 | 0 | 0 |
| Manage / MG.4.1 | 8 | 8 | 0 | 0 |
| Manage / MG.4.2 | 4 | 4 | 0 | 0 |
| Measure / MS.2.7 | 4 | 4 | 0 | 0 |
| Measure / MS.3.3 | 4 | 4 | 0 | 0 |
ISO/IEC 42001:2023
Rules with at least one iso_42001 mapping: 33
| Reference | Rules | Primary | Secondary | Partial |
| Clause 8.4 | 33 | 33 | 0 | 0 |
Per-Rule Detail
Each entry includes the rule SHA-256 (tamper-evidence), provenance, and
full compliance context. Rules sorted by ID for stable diffs.
ATR-2026-10882 — Screen Capture Activity Via Psr.EXE medium
SHA-256: 79694c5995b45ca022067d3a51f79505d8a4ffd45900d44b560b06d5ee7691be
Severity: medium · Source: sigma 2158f96f-43c2-43cb-952a-ab4580f32382 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — screen capture by agent leaks UI content not in agent context.
- Article 50 (primary) — Article 50 transparency — users should know agent is recording screen.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — psr.exe used for surveillance.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure via screen content.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime detection of unsanctioned data capture.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for surveillance via agent.
ATR-2026-12808 — Process Initiated Network Connection To Ngrok Domain high
SHA-256: ea5c0bf6817102eb0eb6de383ac93d64aabb69931309c80da8c54189c8165ca0
Severity: high · Source: sigma 18249279-932f-45e2-b37a-8925f2597670 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — agents fetching from or posting to ngrok tunnels are operating outside their declared scope. Often correlated with prompt-injection that exfiltrates context_window contents.
- Article 12 (primary) — Article 12 logging — every ngrok-bound tool_call should be logged for post-incident traceability. Firing this rule is the logging trigger.
OWASP Agentic Top 10 (2026)
- ASI02:2026 (primary) — ASI02:2026 Memory/Context Manipulation — exfiltrating context_window or memory_access contents via outbound HTTP to attacker tunnel.
- ASI05:2026 (primary) — ASI05:2026 Tool Misuse — using legitimate HTTP tools for exfiltration to non-business destinations.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — LLM02:2025 Sensitive Information Disclosure — exfiltration tunnels are the canonical disclosure channel.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — NIST AI RMF MG.2.3 — runtime treatment of identified data-exfiltration risks. Detection at the egress point is the primary control.
ATR-2026-13292 — Sticky Key Like Backdoor Execution critical
SHA-256: eea69e5131c577f5d9d7f4083af251a5cce2cce030973fd407d054edabeee0c5
Severity: critical · Source: sigma 2fdefcb3-dbda-401e-ae23-f0db027628bc · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 — backdoor install requires absolute human approval.
- Article 15 (primary) — Article 15 robustness — backdoor enables auth bypass.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise — sticky-keys gives SYSTEM at lock screen.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent installing auth backdoor.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-13963 — Triple Cross eBPF Rootkit Execve Hijack high
SHA-256: 66db1c35b971d3c074a6753d79c4c1dd57ff9432e7b1349277eb1901a2621ef4
Severity: high · Source: sigma 0326c3c8-7803-4a0f-8c5c-368f747f7c3e · Rule version: 1 · Detection field: Image
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging applies generically: if the host running the agent process is compromised by a kernel rootkit, the agent's audit log integrity is at risk. The rule belongs in defense-in-depth host monitoring, not in agent runtime.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — NIST AI RMF GV.1.1 (governance policies) calls for host-environment integrity as a precondition for trustworthy AI. This rule supports the host-integrity precondition — it is not an AI-specific control.
ATR-2026-14122 — User Has Been Deleted Via Userdel medium
SHA-256: d5400aa4b801995c157679b819cd5bade9176f0d3ed7372bdb3e1629c20e315f
Severity: medium · Source: sigma 08f26069-6f80-474b-8d1f-d971c6fedea0 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — irreversible identity actions REQUIRE human approval.
- Article 15 (primary) — Article 15 robustness against destructive cascade.
OWASP Agentic Top 10 (2026)
- ASI07:2026 (primary) — Cascading Tool Misuse — identity removal can cascade across systems.
- ASI06:2026 (secondary) — Privilege Compromise enabling user deletion.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent should never delete users.
NIST AI RMF 1.0
- Manage / MG.4.2 (primary) — Incident response trigger for irreversible identity actions.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for destructive identity ops.
ATR-2026-14813 — Procdump Execution medium
SHA-256: 14b0ddfa88489664a5c10ba913d7efa1d31298d5e6bfd06914980896a1e50f3a
Severity: medium · Source: sigma 2e65275c-8288-4ab4-aeb7-6274f58b6b20 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 — lsass dump requires human gate; never agent-initiated.
- Article 15 (primary) — Article 15 robustness — credential dump via agent.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise — lsass dump yields domain credentials.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure — credentials in memory dump.
NIST AI RMF 1.0
- Manage / MG.4.2 (primary) — Incident response trigger.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-16162 — Touch Suspicious Service File medium
SHA-256: 95b3386f30b334ad413a87e353c6377f81815e5c81cf8060863d82051b459132
Severity: medium · Source: sigma 31545105-3444-4584-bebf-c466353230d2 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — persistent root service install requires approval.
- Article 15 (primary) — Article 15 robustness against persistence.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via persistent root service.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent installing systemd persistence.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-17154 — Remote Access Tool - Potential MeshAgent Execution - MacOS medium
SHA-256: d326f9da9fd24f1b6b57774f2a05a087fdf3b4b6582c25c12814c13681d5cf7a
Severity: medium · Source: sigma 22c45af6-f590-4d44-bab3-b5b2d2a2b6d9 · Rule version: 1 · Detection field: CommandLine
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — host RAT presence affects AI integrity.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host integrity.
ATR-2026-20444 — UAC Bypass Using Consent and Comctl32 - Process high
SHA-256: e5e9e6d8bb88956965ea1a1c0d1dab8a75be3bb73f6ef3084f6e619a87391981
Severity: high · Source: sigma 1ca6bd18-0ba0-44ca-851c-92ed89a61085 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — UAC bypass defeats the very gate Microsoft built for human approval.
- Article 15 (primary) — Article 15 robustness — agent should not have privilege-elevation capability.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise — UAC bypass elevates agent to admin.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent self-elevating privileges.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent altering its own privileges.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for UAC bypass via AI agent.
ATR-2026-20705 — Remote Access Tool - Potential MeshAgent Execution - Windows medium
SHA-256: b4f69a2cbfc6a5add1cc6d3243026fd3369402732aa9fff321ead5d2d817eac9
Severity: medium · Source: sigma 2fbbe9ff-0afc-470b-bdc0-592198339968 · Rule version: 1 · Detection field: CommandLine
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — host RAT presence.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host integrity.
ATR-2026-21109 — Deletion of Volume Shadow Copies via WMI with PowerShell high
SHA-256: 408b1e1d59ae285a1da0037f21021ddc4826ae88067e00d532bf52edc4657abf
Severity: high · Source: sigma 21ff4ca9-f13a-41ad-b828-0077b2af2e40 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 human oversight is REQUIRED for irreversible high-impact actions. Shadow-copy deletion is irreversible (no recovery without backups). Any agent performing this without explicit human approval fails Article 14 compliance.
- Article 15 (primary) — Article 15 robustness — destructive cascade resistance. Agents should refuse or escalate destructive actions; firing this rule is direct evidence the safeguard failed.
- Article 9 (primary) — Article 9 risk management — destructive data actions must be in the risk register with active controls. This rule IS the active control.
OWASP Agentic Top 10 (2026)
- ASI07:2026 (primary) — ASI07:2026 (Cascading Tool Misuse) — agent chains read+delete operations to make data recovery impossible.
- ASI06:2026 (secondary) — Privilege compromise enabling the destructive action.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — LLM06:2025 Excessive Agency — destructive admin commands are the textbook example.
NIST AI RMF 1.0
- Manage / MG.4.2 (primary) — NIST AI RMF MG.4.2 — incident response capability for AI-driven harm. Shadow-copy deletion via agent is a high-severity incident requiring immediate playbook execution.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for destructive AI actions — irreversibility makes this a top-priority assessment item.
ATR-2026-24756 — Time Travel Debugging Utility Usage high
SHA-256: 30fac93824b5ca59129e6ac9381c0eaa743ab38d85d9565194b4e56195798ea5
Severity: high · Source: sigma 0b4ae027-2a2d-4b93-8c7e-962caaba5b2a · Rule version: 1 · Detection field: ParentImage
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging applies generically; agent-host debugger compromise affects AI integrity.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host-environment debugger access boundary.
ATR-2026-26270 — Potential Provisioning Registry Key Abuse For Binary Proxy Execution high
SHA-256: d67c22a575f2a6f5ee9e2bdd184b89eceba867b0a9db1c581d264bcd540cc839
Severity: high · Source: sigma 2a4b3e61-9d22-4e4a-b60f-6e8f0cde6f25 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — provisioning install bypasses standard install gates.
- Article 15 (primary) — Article 15 robustness against persistence install via provisioning.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via provisioning persistence.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency for system-level provisioning install.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent altering system provisioning.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for provisioning abuse.
ATR-2026-27041 — OMIGOD SCX RunAsProvider ExecuteShellCommand high
SHA-256: 498061fa02ffdf2d3f44ae384ce691c7951f0c08dd5cde3d4df30e45b9685768
Severity: high · Source: sigma 21541900-27a9-4454-9c4c-3f0a4240344a · Rule version: 1 · Detection field: User
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — agent host running on Azure with OMI may be compromised.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host-environment integrity.
ATR-2026-31863 — Suspicious Processes Spawned by Java.EXE high
SHA-256: ff4b6da7668d4fab58d3886140fa5065229ad8bfdfb8a3ce8f8b603299ebcd30
Severity: high · Source: sigma 0d34ed8b-1c12-4ff2-828c-16fc860b766d · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — Java deserialization / Log4Shell-style RCE via agent.
- Article 9 (primary) — Article 9 risk management — Log4Shell-class CVEs are documented AI deployment risks.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via Java RCE.
OWASP LLM Top 10 (2025)
- LLM03:2025 (primary) — Supply Chain — vulnerable Java dependency in agent runtime.
NIST AI RMF 1.0
- Manage / MG.3.2 (primary) — Managing third-party dependency risk in AI runtime.
ATR-2026-34130 — Disk Image Creation Via Hdiutil - MacOS medium
SHA-256: dbaea04ddbc580dffc98c9a4c030c01a2d85f4e0ad2df88da5b367b1630d3812
Severity: medium · Source: sigma 1cf98dc2-fcb0-47c9-8aea-654c9284d1ae · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent staging payload via DMG creation/mount.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — hdiutil for payload packaging.
OWASP LLM Top 10 (2025)
- LLM06:2025 (secondary) — Excessive Agency — agent should not mount disk images at runtime.
NIST AI RMF 1.0
- Measure / MS.3.3 (primary) — Detect agent operating outside expected scope.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for macOS payload-staging.
ATR-2026-34505 — Remote File Download Via Desktopimgdownldr Utility medium
SHA-256: ed0196324cd5193a9f1ff159260f1dc20fc45998bf67d0b7ffef39c822a342d9
Severity: medium · Source: sigma 214641c2-c579-4ecb-8427-0cf19df6842e · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent using OS LOLBIN to fetch second-stage payload.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — desktopimgdownldr repurposed as downloader.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent should not have arbitrary download.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of LOLBIN download.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-36892 — Base64 MZ Header In CommandLine high
SHA-256: a28161435ffadd8f32613dce0b6a1470bac1d80e0ebae56dc5cf7b6728c5ba48
Severity: high · Source: sigma 22e58743-4ac8-4a9f-bf19-00a0428d8c5f · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — embedded PE payload via agent shell or prompt is direct exploit attempt.
- Article 9 (primary) — Article 9 risk management — embedded-payload exploitation is documented attack class.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via in-memory PE load.
- ASI01:2026 (secondary) — Goal Hijack — payload injected via user_input vector.
OWASP LLM Top 10 (2025)
- LLM01:2025 (primary) — Prompt Injection embedding base64 PE payload.
- LLM06:2025 (primary) — Excessive Agency executing decoded payload.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of payload-injection risk.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for payload exec via agent.
ATR-2026-38740 — Access of Sudoers File Content medium
SHA-256: 75b37715aea519a4f06c5e0804dfc1973d3e211270a2d941e0f70961705dd464
Severity: medium · Source: sigma 0f79c4d2-4e1f-4683-9c36-b5469a665e06 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — reading sudoers is privilege-escalation reconnaissance.
- Article 50 (secondary) — Article 50 transparency — agent reading sensitive files should be logged for users.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — shell access used to read privileged files.
- ASI06:2026 (primary) — Privilege Compromise reconnaissance.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure — sudoers contents can be exfiltrated.
NIST AI RMF 1.0
- Measure / MS.3.3 (primary) — Detect agent operating outside expected file-access scope.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for privileged-file access via agent.
ATR-2026-39365 — Suspicious X509Enrollment - Process Creation medium
SHA-256: 9411a0fa9c204a6fcd81dfa92aa54cf5784052a641d0ace70a867739ab7cc9af
Severity: medium · Source: sigma 114de787-4eb2-48cc-abdb-c0b449f93ea4 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — installing rogue trust roots breaks the entire TLS/code-signing trust chain agent runtime depends on.
- Article 14 (primary) — Article 14 oversight — trust-root install is high-impact and requires human approval.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise — rogue cert enables persistent MITM.
- ASI09:2026 (primary) — Identity Spoofing — rogue cert impersonates legitimate authority.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — cert install should never be agent-initiated.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent altering trust boundaries.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for trust-anchor manipulation via AI agent.
ATR-2026-41443 — Linux Recon Indicators high
SHA-256: 77eaf912172a9294eb316e19bb7273b2904074a830464621e4a12c33e59f83e9
Severity: high · Source: sigma 0cf7a157-8879-41a2-8f55-388dd23746b7 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness covers detection of agent steering toward unauthorized data discovery. SUID-binary and credential-file enumeration is reconnaissance preparation for privilege escalation; firing this rule is direct evidence of an agent acting outside its intended scope.
- Article 50 (secondary) — Article 50 (transparency) — when agent is doing reconnaissance, end users should be alerted; the rule provides the trigger.
OWASP Agentic Top 10 (2026)
- ASI04:2026 (primary) — ASI04:2026 (Resource Overload) and ASI05:2026 (Tool Misuse) — using shell tools for reconnaissance is canonical tool misuse beyond the agent task scope.
NIST AI RMF 1.0
- Measure / MS.3.3 (primary) — Measuring agent behavior against expected scope — recon commands are a clear out-of-scope signal that NIST MS.3.3 monitoring should capture.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for AI deployment must account for agent-driven reconnaissance as a misuse vector.
ATR-2026-43105 — Start of NT Virtual DOS Machine medium
SHA-256: 4ff1a85698ec11ff114aacc4b3b98a7e5faa44dd31c255ce94fedc3c9c4b54ac
Severity: medium · Source: sigma 16905e21-66ee-42fe-b256-1318ada2d770 · Rule version: 1 · Detection field: Image
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging applies generically; NTVDM activity on agent host is host-EDR concern.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host-environment integrity precondition.
ATR-2026-49108 — Use of Pcalua For Execution medium
SHA-256: 8c1d07906f1c58a9e5bd2dcaadf370e1e5dec49ac9d66558fb657f7ad6387160
Severity: medium · Source: sigma 0955e4e1-c281-4fb9-9ee1-5ee7b4b754d2 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — LOLBIN evasion via agent shell.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — pcalua proxy execution.
NIST AI RMF 1.0
- Measure / MS.2.7 (primary) — LOLBIN technique detection.
ATR-2026-50378 — HackTool - CrackMapExec Execution Patterns high
SHA-256: c46f5b1301c78ed6b4b22070608b7498f5212233e2b67d3a82a8d9aabb182655
Severity: high · Source: sigma 058f4380-962d-40a5-afce-50207d36d7e2 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — agent issuing AD enumeration is direct evidence of out-of-scope reconnaissance.
- Article 12 (primary) — Article 12 logging — every AD-enum tool_call must be auditable.
OWASP Agentic Top 10 (2026)
- ASI04:2026 (primary) — Resource Overload — bulk AD enumeration via agent.
- ASI06:2026 (primary) — Privilege Compromise — AD recon precedes lateral movement.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency for post-exploit framework usage.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of agent-driven AD reconnaissance risk.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for AD reconnaissance via AI agent.
ATR-2026-50890 — JAMF MDM Potential Suspicious Child Process medium
SHA-256: 61b0b4890b6bb4e2b230e2e977deecf462bc68e38c8d31c73b71ddcaedd4a170
Severity: medium · Source: sigma 2316929c-01aa-438c-970f-099145ab1ee6 · Rule version: 1 · Detection field: ParentImage
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — MDM compromise of agent host.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host integrity.
ATR-2026-53865 — Network Communication Initiated To Portmap.IO Domain medium
SHA-256: 148607c85a51b09a7f09866287c172c7ff7833a8a54d672356d7fcfcab2187a7
Severity: medium · Source: sigma 07837ab9-60e1-481f-a74d-c31fb496a94c · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — tunneling-service connections from agent indicate scope violation.
- Article 12 (primary) — Article 12 logging requirements for outbound C2-class destinations.
OWASP Agentic Top 10 (2026)
- ASI02:2026 (primary) — Memory/Context Manipulation — context exfiltration to attacker-controlled tunnel.
- ASI05:2026 (primary) — Tool Misuse — HTTP tools used for non-business destinations.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure via tunnel.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime data-exfiltration detection at egress.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for outbound tunneling via AI agent.
ATR-2026-54358 — Shell Execution Of Process Located In Tmp Directory high
SHA-256: 123eed966e55556ebad65f01441dc2ad5bcecac95fff8470c7ea9024a2428c59
Severity: high · Source: sigma 2fade0b6-7423-4835-9d4f-335b39b83867 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent executing binary from tmp is dropper-pattern signal.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — agent shell launching dropper binary.
OWASP LLM Top 10 (2025)
- LLM06:2025 (secondary) — Excessive Agency.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime detection of dropper staging.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-64888 — MacOS Scripting Interpreter AppleScript medium
SHA-256: 161fd7dfeabe6cca06760c343036de7098ff1c97958525fc5646659ef741ed7c
Severity: medium · Source: sigma 1bc2e6c5-0885-472b-bed6-be5ea8eace55 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent invoking osascript for keychain access or credential dialog.
- Article 50 (primary) — Article 50 transparency — credential prompts via agent need user disclosure.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — scripting interpreter used for credential theft.
- ASI09:2026 (primary) — Identity Spoofing — fake credential prompt impersonates system.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure — credentials harvested.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime credential-theft detection.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for AppleScript abuse via agent.
ATR-2026-65860 — Suspicious Splwow64 Without Params high
SHA-256: fb3542dc5ab808534a576dfc327e025bfd52e4a876f59b07af44529c33371d75
Severity: high · Source: sigma 1f1a8509-2cbb-44f5-8751-8e1571518ce2 · Rule version: 1 · Detection field: Image
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — host-EDR concern.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host-environment integrity.
ATR-2026-70113 — Communication To Ngrok Tunneling Service Initiated high
SHA-256: 79a10fba5af1bfd35cdf0f092c7511060e528aaa50356bc5e6fea4c56079c2d3
Severity: high · Source: sigma 1d08ac94-400d-4469-a82f-daee9a908849 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — outbound to tunneling service.
OWASP Agentic Top 10 (2026)
- ASI02:2026 (primary) — Memory/Context Manipulation via tunnel exfil.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure via tunnel.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime data-exfiltration detection.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for outbound tunneling.
ATR-2026-70804 — Uncommon Network Connection Initiated By Certutil.EXE high
SHA-256: 9d75a07d4a422a685cfb834c6473b05d95b56b4bbf23a786ce1b4a82728436de
Severity: high · Source: sigma 0dba975d-a193-4ed1-a067-424df57570d1 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — agent using OS-built-in download cradles to fetch second-stage payloads.
- Article 12 (primary) — Article 12 logging — every certutil network invocation needs audit trail.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — certutil repurposed as a download tool by agent.
- ASI06:2026 (secondary) — Privilege Compromise via second-stage staging.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent should not have arbitrary download capability.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of LOLBIN download via agent.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for agent-driven payload staging.
ATR-2026-71531 — Suspicious Outlook Child Process high
SHA-256: d5be401e4b0c760e3e5dfa3aedb7ead5a58710c8b8110a336fa455fab277be3f
Severity: high · Source: sigma 208748f7-881d-47ac-a29c-07ea84bf691d · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — email agents resolving attachments are common phishing vector.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — email tools chained to shell exec.
OWASP LLM Top 10 (2025)
- LLM01:2025 (primary) — Prompt Injection — embedded macros / HTML in email content.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of email-driven exec.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for email-handling agent.
ATR-2026-73413 — Suspicious HWP Sub Processes high
SHA-256: 89543e1a86a2d6d7ed3e4e36d6d70aca1ed008942af5b1ef425e108c478e05b8
Severity: high · Source: sigma 023394c4-29d5-46ab-92b8-6a534c6f447b · Rule version: 1 · Detection field: ParentImage
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging applies generically; host-document compromise of the agent operator host can affect AI integrity.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host-environment integrity precondition for trustworthy AI deployment.
ATR-2026-74985 — Suspicious Chromium Browser Instance Executed With Custom Extension high
SHA-256: 566b938393d52e77b2f4042d3e06e230f4836f45cbe6a8f5548118acf0c91701
Severity: high · Source: sigma 27ba3207-dd30-4812-abbf-5d20c57d474e · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 cybersecurity — loading a third-party extension into the agent's browser context expands the trust boundary unpredictably. Robustness violation when the extension is attacker-controlled.
- Article 14 (primary) — Article 14 oversight — extension installation should require human gating, not be performed via tool_call directive.
OWASP Agentic Top 10 (2026)
- ASI08:2026 (primary) — ASI08:2026 Skill/Plugin Compromise — loading arbitrary browser extensions through tool_call is the browser-context analogue of skill compromise.
OWASP LLM Top 10 (2025)
- LLM03:2025 (primary) — LLM03:2025 Supply Chain — third-party extension is a supply-chain trust boundary violation when loaded mid-session.
NIST AI RMF 1.0
- Manage / MG.3.2 (primary) — NIST AI RMF MG.3.2 — managing third-party tooling supply chain. Browser extensions are third-party plugins.
ATR-2026-75586 — Dumping Process via Sqldumper.exe medium
SHA-256: e0d9dcd162bf7d2eae61854e39df40a19f84e9374a4aa56e440fb87ca7e7f16e
Severity: medium · Source: sigma 23ceaf5c-b6f1-4a32-8559-f2ff734be516 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent dumping process memory exfils credentials.
- Article 12 (primary) — Article 12 logging — every memory-dump tool_call needs audit.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — sqldumper repurposed as credential dumper.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure — memory dump contains credentials.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime credential-theft detection.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for memory-dump via agent.
ATR-2026-76233 — HackTool - Wmiexec Default Powershell Command high
SHA-256: 9b9a4d5d80467a57640a053ad9a276422a10e304d68aee80584358bd4246431b
Severity: high · Source: sigma 022eaba8-f0bf-4dd9-9217-4604b0bb3bb0 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — lateral-movement via agent shell tool defeats single-host AI deployment isolation assumptions.
- Article 14 (primary) — Article 14 oversight — privileged remote-exec actions need human-in-the-loop gating.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via lateral movement using delegated shell tool authority.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent should not have remote-exec capability for typical use cases.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of lateral-movement risk for AI agents with shell access.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — AI deployment impact assessment must account for lateral-movement misuse.
ATR-2026-77355 — UtilityFunctions.ps1 Proxy Dll medium
SHA-256: 547c3aacc43b4a13cc35820d88f5e4f358bdff7c9c6b004b815c077abd10dde4
Severity: medium · Source: sigma 0403d67d-6227-4ea8-8145-4e72db7da120 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — proxy-DLL helper signals defense-evasion attempt via agent shell.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via DLL hijack helper functions.
OWASP LLM Top 10 (2025)
- LLM06:2025 (secondary) — Excessive Agency — agent should not load proxy-DLL helpers.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of agent-driven defense evasion.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for agent-driven DLL hijack.
ATR-2026-77646 — Potential Persistence Via Logon Scripts - CommandLine high
SHA-256: fd134e4ba30395188b8195c3d1a8fae235cee5a8ce9d4deb9cbd37f00c1aeeac
Severity: high · Source: sigma 21d856f9-9281-4ded-9377-51a1a6e2a432 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — persistence install requires approval.
- Article 15 (primary) — Article 15 robustness against persistence install via agent.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via persistence at logon.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency for persistence install.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent installing persistence.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-77772 — Potential Defense Evasion Activity Via Emoji Usage In CommandLine - 4 high
SHA-256: e4e5832432a2714c4484d61fb6245320d838f9c3377b75ac2a687abf6b3a2280
Severity: high · Source: sigma 225274c4-8dd1-40db-9e09-71dff4f6fb3c · Rule version: 1 · Detection field: user_input
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness against adversarial input. Emoji-based obfuscation is a known evasion technique against keyword filters; detecting it is part of robustness assurance.
OWASP Agentic Top 10 (2026)
- ASI01:2026 (primary) — Agent Goal Hijack — obfuscated payloads in user_input attempt to override agent instructions while evading detection.
OWASP LLM Top 10 (2025)
- LLM01:2025 (primary) — LLM01:2025 Prompt Injection — emoji obfuscation is documented in the OWASP prompt-injection taxonomy.
NIST AI RMF 1.0
- Measure / MS.2.7 (primary) — Measuring AI system security against documented adversarial techniques. Emoji obfuscation is a published technique requiring active detection.
ATR-2026-78212 — Potential Hidden Directory Creation Via NTFS INDEX_ALLOCATION Stream - CLI medium
SHA-256: 6c5536e4c8ac24e3da20307ecdeb9b58160f1ec6cfa0a1c32f72ea2939772fba
Severity: medium · Source: sigma 0900463c-b33b-49a8-be1d-552a3b553dae · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — defense-evasion technique attempt via agent shell.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — using filesystem tools for evasion.
NIST AI RMF 1.0
- Measure / MS.3.3 (primary) — Detect agent operating outside expected scope.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for filesystem-evasion via agent.
ATR-2026-79114 — Hacktool Execution - Imphash critical
SHA-256: c2676442f8ebadc58fd835edcc3ae5ec234dd989c9a0cbdd4939a43357519543
Severity: critical · Source: sigma 24e3e58a-646b-4b50-adef-02ef935b9fc8 · Rule version: 1 · Detection field: Hashes
AI agent analogue: NO · LLM-enriched: YES
Needs human review (2 items):
- detection.conditions[].field (endpoint→agent-context mapping)
- detection.conditions[].field (LLM determined no agent analogue exists; rule kept with original endpoint fields and will not activate against agent events)
EU AI Act
- Article 12 (partial) — Article 12 logging — host hacktool execution affects agent host integrity.
NIST AI RMF 1.0
- Govern / GV.1.1 (partial) — Host integrity precondition.
ATR-2026-81735 — HackTool - SILENTTRINITY Stager Execution high
SHA-256: e1ae2a2369a9062ef2928aa589d95c43acce7fdb242d982ba0b9a362ecd55e56
Severity: high · Source: sigma 03552375-cc2c-4883-bbe4-7958d5a980be · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — reflective code loading via agent shell defeats sandboxing.
- Article 12 (primary) — Article 12 logging — every reflective-load tool_call needs audit trail.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via in-memory code execution.
- ASI07:2026 (secondary) — Cascading Tool Misuse — chain of download + decode + execute.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — Excessive Agency — agent should not have arbitrary code execution.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — Managing risk of agent modifying its own runtime via reflective loading.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for in-process code injection via AI agent.
ATR-2026-83241 — Network Connection Initiated via Finger.EXE high
SHA-256: 258c71b2af9ffc72ea0914d5f2222eefab6489564406b85f211f6bb41a6628c6
Severity: high · Source: sigma 2fdaf50b-9fd5-449f-ba69-f17248119af6 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — finger as covert C2 channel via agent.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — finger repurposed as C2.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Information disclosure via finger response.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime egress detection on uncommon protocols.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment.
ATR-2026-85501 — Malicious PowerShell Commandlets - ProcessCreation high
SHA-256: 736dbd52c77642d8befa93d1c55ba7670ab2cc73568e4b2ff83eef39d07ee740
Severity: high · Source: sigma 02030f2f-6199-49ec-b258-ea71b07e03dc · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 requires high-risk AI systems to be resilient against malicious manipulation. An agent invoking PowerSharpPack/PowerView cmdlets via tool_call.arguments is direct evidence the agent has been steered toward attacker objectives — a robustness failure that this rule surfaces in real time.
- Article 12 (secondary) — Article 12 mandates automatic logging of high-risk AI events. Triggering on this rule produces the audit-grade event needed to demonstrate post-incident traceability of agent compromise.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — OWASP Agentic ASI06:2026 (Privilege Compromise) covers exactly this class — agent with delegated authority is induced to run privileged tooling. The cmdlets in the rule pattern (Invoke-Mimikatz, DCSync) are the canonical post-compromise toolset.
- ASI01:2026 (secondary) — When the agent's goal is hijacked into running these cmdlets, ASI01 (Agent Goal Hijack) is the upstream vector that the rule indirectly reveals.
OWASP LLM Top 10 (2025)
- LLM06:2025 (partial) — LLM06:2025 (Excessive Agency) — the agent has tools that can execute arbitrary cmdlets and is induced to misuse them. The rule fires on the misuse, not the agency itself, so 'partial'.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — NIST AI RMF MG.2.3 calls for active runtime countermeasures against identified agent risks. Detection of exploitation cmdlet invocation in tool_call arguments is the primary risk-treatment control for the Privilege Compromise risk class.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — ISO 42001 clause 8.4 requires impact assessment of AI deployments to consider adversarial misuse. Detection events from this rule are the monitoring evidence the impact assessment process needs.
ATR-2026-88829 — Powershell Defender Exclusion medium
SHA-256: 5590aeb152f707f051d59e6db37b76d37a0b542ad8819e9f7f214b9fcbebdb13
Severity: medium · Source: sigma 17769c90-230e-488b-a463-e05c08e9d48f · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — disabling host AV via agent action degrades the integrity of the AI deployment environment. Direct robustness violation.
- Article 14 (primary) — Article 14 (human oversight) — adding AV exclusions is a high-impact action that should require human-in-the-loop approval. Firing this rule means the action happened without that gate.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — ASI06:2026 Privilege Compromise — agent leveraging admin privileges to weaken host security posture. Canonical example.
- ASI09:2026 (secondary) — ASI09:2026 (Identity Spoofing) — when an agent runs as a privileged service identity, AV-exclusion actions inherit that identity and circumvent oversight.
OWASP LLM Top 10 (2025)
- LLM06:2025 (primary) — LLM06:2025 Excessive Agency — agent with privileges to modify security software is excessive for typical use cases. The rule surfaces concrete misuse.
NIST AI RMF 1.0
- Manage / MG.4.1 (primary) — NIST AI RMF MG.4.1 — managing risks of AI systems modifying their operating environment. AV exclusion is the highest-impact form.
ATR-2026-90343 — Communication To Ngrok Tunneling Service - Linux high
SHA-256: 6aebe61cf64e5d0ad9b3a4b4ab62ec7008447f1736e819b017365ec6122b53e6
Severity: high · Source: sigma 19bf6fdb-7721-4f3d-867f-53467f6a5db6 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — Linux agent reaching ngrok tunnel signals scope violation.
- Article 12 (primary) — Article 12 logging — outbound C2 destinations.
OWASP Agentic Top 10 (2026)
- ASI02:2026 (primary) — Memory/Context Manipulation — context exfil to attacker tunnel.
- ASI05:2026 (primary) — Tool Misuse — HTTP tools used for non-business destinations.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure via tunnel.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime treatment of data-exfiltration risk.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for outbound tunneling via Linux agent.
ATR-2026-90471 — Shell Execution via Nice - Linux high
SHA-256: 75e168288571639e30df70ff5ee767c0929bcbfa5dc0d9cb9823f905e7337045
Severity: high · Source: sigma 093d68c7-762a-42f4-9f46-95e79142571a · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 robustness — LOLBIN evasion via agent shell.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — using nice as a shell wrapper to evade detection.
NIST AI RMF 1.0
- Measure / MS.2.7 (primary) — Measuring against documented evasion techniques.
ATR-2026-93338 — Suspicious Double Extension File Execution high
SHA-256: a3fe9f4b4209169aff53b73238384a29f07a22d8fd148853098358d8df236eae
Severity: high · Source: sigma 1cdd9a09-06c9-4769-99ff-626e2b3991b8 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent resolving double-extension file is documented social-engineering vector.
OWASP Agentic Top 10 (2026)
- ASI05:2026 (primary) — Tool Misuse — file tools used to launch deceptively-named executable.
OWASP LLM Top 10 (2025)
- LLM01:2025 (secondary) — Prompt Injection often instructs agent to "open the document".
NIST AI RMF 1.0
- Measure / MS.2.7 (primary) — Measure against documented social-engineering techniques.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for double-extension exec via agent.
ATR-2026-94647 — Invocation of Active Directory Diagnostic Tool (ntdsutil.exe) medium
SHA-256: fd278b820457a5f941774246e84c2108b79d21c080733b5e8b4d971676621198
Severity: medium · Source: sigma 2afafd61-6aae-4df4-baed-139fa1f4c345 · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 14 (primary) — Article 14 oversight — AD password database dump is the highest-impact action; absolute requirement for human approval.
- Article 15 (primary) — Article 15 robustness — agent dumping NTDS.dit is direct critical-asset compromise.
- Article 9 (primary) — Article 9 risk management — NTDS dump is documented top-tier AI deployment risk.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise — full domain compromise via NTDS dump.
- ASI07:2026 (primary) — Cascading Tool Misuse — single dump compromises every account.
OWASP LLM Top 10 (2025)
- LLM02:2025 (primary) — Sensitive Information Disclosure at scale (every domain credential).
NIST AI RMF 1.0
- Manage / MG.4.2 (primary) — Incident response — NTDS dump triggers IR playbook immediately.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for AD-database dump via agent.
ATR-2026-97982 — Potentially Suspicious WebDAV LNK Execution medium
SHA-256: a15113d62a763b740eb72f72d52efd919cfbe193fc9723b413ae6f6c6047665c
Severity: medium · Source: sigma 1412aa78-a24c-4abd-83df-767dfb2c5bbe · Rule version: 1 · Detection field: tool_call.arguments
AI agent analogue: YES · LLM-enriched: YES
EU AI Act
- Article 15 (primary) — Article 15 — agent invoking remote-payload LNK from WebDAV defeats local-execution boundary.
OWASP Agentic Top 10 (2026)
- ASI06:2026 (primary) — Privilege Compromise via remote payload load.
OWASP LLM Top 10 (2025)
- LLM03:2025 (primary) — Supply Chain — remote payload is a supply-chain trust violation.
NIST AI RMF 1.0
- Manage / MG.2.3 (primary) — Runtime detection of remote payload load via agent.
ISO/IEC 42001:2023
- Clause 8.4 (primary) — Impact assessment for remote-payload execution.
Honest Disclosures
This pack documents migrator-converted detection rules derived
from upstream Sigma/YARA sources. Conversion fidelity varies per rule:
- 40 rules are LLM-enriched
and reauthored with detection fields appropriate for AI agent runtime
context (
tool_call.arguments, user_input,
agent_output, etc).
- 10 rules are auto-generated
placeholders that retain the original Sigma/YARA endpoint detection
fields. These rules are schema-valid and can be deployed, but will only
fire on host OS event telemetry, not on agent runtime events. Each
carries a
needs_human_review marker explicitly listing the
review items.
The migrator deliberately preserves rules with no agent analogue
(has_agent_analogue: false) rather than fabricate translations.
These rules document the operator's detection coverage but are not
expected to activate against agent telemetry.
For full audit-grade detection coverage, replace placeholder rules with
hand-crafted enrichments (LLM-assisted or manual review by a security
engineer).
Auditor signature: __________________________________
Date: ____________________