OWASP_LLM_TOP10LLM06:2025voice-validated

OWASP_LLM_TOP10 LLM06: LLM06:2025

OWASP_LLM_TOP10

AL
Adam Lundqvist
Founder at SQUR · last verified 2026-06-20

Regulation text

Excessive agency arises when LLM-based systems are granted excessive functionality, permissions, or autonomy. Damaging actions can occur in response to unexpected or ambiguous outputs from the LLM, regardless of root cause. Includes over-permissive tools/plugins, autonomous action without human confirmation for high-impact operations, and lack of fine-grained authorisation on agent-callable functions.

ATT&CK techniques this article tests · 15

TechniqueWhy it mapsConfidence
T1059.0041. LLM with excessive agency can execute arbitrary Unix shell commands, leading to system compromise. This directly exploits over-permissive tool access, as described in LLM06:2025.
90%
T1059.0032. LLM with excessive agency can execute arbitrary Windows command shell commands, enabling system control. This stems from insufficient authorization for agent-callable functions, as per LLM06:2025.
90%
T1059.0013. LLM with excessive agency can execute arbitrary PowerShell commands, facilitating system manipulation. This is due to over-permissive tools or plugins, a key concern in LLM06:2025.
90%
T14854. LLM with excessive agency can delete critical data, causing irreversible damage. Autonomous action without human confirmation enables this impact, as highlighted in LLM06:2025.
80%
T14905. LLM with excessive agency can inhibit system recovery by deleting backups or shadow copies. Lack of fine-grained authorization allows such destructive actions, as per LLM06:2025.
80%
T14866. LLM with excessive agency can encrypt data for impact, leading to data unavailability. Over-permissive tools enable this ransomware-like behavior, a risk identified in LLM06:2025.
70%
T15317. LLM with excessive agency can remove user accounts, disrupting operations and access. Autonomous action without human confirmation facilitates this, as described in LLM06:2025.
70%
T10688. LLM with excessive agency can exploit system vulnerabilities to escalate privileges. Over-permissive functionality allows the LLM to attempt such exploits, as per LLM06:2025.
70%
T15629. LLM with excessive agency can impair defenses by disabling security software. Lack of fine-grained authorization on agent-callable functions enables this, as noted in LLM06:2025.
80%
T1070.00410. LLM with excessive agency can delete logs and other indicators, hindering incident response. Over-permissive tools allow for defense evasion, a concern in LLM06:2025.
70%
T100511. LLM with excessive agency can collect sensitive data from local systems. This is a direct consequence of over-permissive access to system resources, as per LLM06:2025.
80%
T103912. LLM with excessive agency can collect sensitive data from network shared drives. Lack of fine-grained authorization allows unauthorized network access, as described in LLM06:2025.
70%
T1071.00113. LLM with excessive agency can establish command and control via web protocols. Over-permissive tools enable unauthorized external communication, a risk in LLM06:2025.
70%
T113314. LLM with excessive agency can provision external remote services, creating new access points. Autonomous action without human confirmation facilitates this, as per LLM06:2025.
60%
T1547.00115. LLM with excessive agency can establish persistence by modifying autostart mechanisms. Over-permissive functionality allows for long-term unauthorized access, as noted in LLM06:2025.
60%

Defending mitigations · 6

MitigationWhat it doesConfidence
M10381. Implement least privilege for LLM agents, ensuring they operate with minimal necessary permissions. This directly counters excessive functionality and autonomy, as per LLM06:2025.
90%
M10262. Strictly control privileged access granted to LLM agents, preventing unauthorized high-impact operations. This addresses over-permissive tools and lack of fine-grained authorization, as per LLM06:2025.
90%
M10303. Isolate LLM environments and restrict network access to limit potential damage. This reduces the scope of excessive agency, as described in LLM06:2025.
80%
M10184. Implement mandatory human review and confirmation for all high-impact actions initiated by the LLM. This prevents autonomous action without human oversight, as highlighted in LLM06:2025.
90%
M10405. Monitor and block unauthorized or anomalous LLM actions in real-time. This prevents damaging actions arising from unexpected LLM outputs, as noted in LLM06:2025.
80%
M10476. Log all LLM actions and tool usage for accountability and detection of excessive agency. This provides crucial audit trails for post-incident analysis, as per LLM06:2025.
80%

Underlying weaknesses · 6

CWEWhy it persistsConfidence
CWE-2691. Improper privilege management directly leads to excessive agency, allowing LLMs to perform unauthorized actions. This is central to the issue described in LLM06:2025.
90%
CWE-2842. Improper access control results in a lack of fine-grained authorization for LLM agent-callable functions. This enables over-permissive tool usage, as per LLM06:2025.
90%
CWE-8623. Missing authorization allows LLMs to execute actions without necessary checks, facilitating autonomous high-impact operations. This is a core aspect of excessive agency, as highlighted in LLM06:2025.
80%
CWE-7324. Incorrect permission assignment for critical resources means tools or functions have excessive power when used by an LLM. This enables over-permissive tools, as described in LLM06:2025.
80%
CWE-2505. Execution with unnecessary privileges allows LLM agents to operate with more power than required, directly contributing to excessive agency. This is a key concern in LLM06:2025.
80%
CWE-6686. Exposure of resources to the wrong sphere makes sensitive tools or functions accessible to the LLM without proper isolation. This enables over-permissive tools/plugins, as noted in LLM06:2025.
70%

What SQUR Covers

Web application + API pentesting for OWASP Top 10, business logic flaws, authentication bypass, injection attacks, and other application-layer vulnerabilities. €1,995 per scan, 24-hour turnaround, EU-only data.

What SQUR Does Not Cover

Internal network pentesting, endpoint security testing, physical security assessments, social engineering, or ICT third-party concentration risk reviews. Engage a complementary provider for those scope items.

Provenance

Mapped Q2.2026 using gemini-2.5-flash · €0.0190 compute · voice-rubric self-validated