⭐ Star on GitHub

Summary: The OWASP Agentic Security Initiative’s threat table — seventeen threats to agentic AI systems, each with a description, mitigation direction, and (where one exists) a mapping back to the OWASP Top 10 for LLM Applications.

Sources: raw/Agentic/Agentic-AI-Threats-and-Mitigations-1.1.pdf

Last updated: 2026-07-29


This is the reference table the rest of the collection hangs off. Threat IDs are stable and used throughout OWASP’s own material, so cite them as T1T17 rather than by name.

The seventeen threats

TIDThreatOne-lineLLM Top 10 relation
T1Memory PoisoningCorrupting short- or long-term memory to alter later decisionsExtends LLM04, LLM08
T2Tool MisuseManipulating an agent into abusing tools it is authorised to usePartially LLM06; also LLM03, LLM08
T3Privilege CompromiseExploiting dynamic role inheritance and permission misconfigurationPartially LLM06
T4Resource OverloadExhausting compute, memory or service quota to degrade the agentRelated to LLM10
T5Cascading Hallucination AttacksFalse output propagating and amplifying through memory and agentsExtends LLM09
T6Intent Breaking & Goal ManipulationRewriting the agent’s planning, reasoning and self-evaluationExtends LLM01
T7Misaligned & Deceptive BehaviorsHarmful action from advanced reasoning, not from error or injection— none —
T8Repudiation & UntraceabilityAgent actions that cannot be traced or attributed— none —
T9Identity Spoofing & ImpersonationImpersonating an agent, user or service; agent-identity theft— none —
T10Overwhelming Human in the LoopExploiting reviewer cognitive limits to force rushed approvals— none —
T11Unexpected RCE and Code AttacksInjecting into AI-generated execution environmentsDistinct from LLM01, LLM05
T12Agent Communication PoisoningManipulating inter-agent channels to corrupt shared knowledgeBeyond LLM04, LLM08
T13Rogue Agents in Multi-Agent SystemsCompromised agents operating outside monitoring boundariesAdversarial exploitation of LLM06
T14Human Attacks on Multi-Agent SystemsExploiting delegation and trust chains to escalate privilege— none —
T15Human ManipulationExploiting user trust in the agent to steer the human— none —
T16Insecure Inter-Agent Protocol AbuseAttacking MCP / A2A: consent bypass, context hijacking— none —
T17Supply Chain CompromisePoisoned prompts, plugins, tool metadata, framework updatesLLM03 amplified

(source: Agentic-AI-Threats-and-Mitigations-1.1.pdf, pp. 16–19)

Seven of seventeen have no Top 10 counterpart

T7, T8, T9, T10, T14, T15 and T16 are named against no OWASP Top 10 entry in the document. They cluster into three groups, and the grouping is the interesting part:

  • Emergent-behaviour threats (T7) — harm arising from the reasoning itself. The document is explicit that this “is distinct from hallucinations, as the behavior emerges from advanced reasoning capabilities rather than just random error or prompt failure,” and concedes the threat “is at an early stage.”
  • Accountability threats (T8, T9) — who did it, and can you prove it. Conventional appsec concerns that the Top 10 never had to state because non-agentic LLM apps do not act. T8 is also STRIDE’s R, a category conventional threat modelling has carried since 1999 — the gap is in the LLM taxonomies, not in threat modelling generally. See threat-modeling-frameworks, and agent-observability-and-logging for what an audit trail has to contain to close it.
  • Human-boundary and protocol threats (T10, T14, T15, T16) — the seams between agents, and between agents and people.

This is the same agentic gap the vault has now recorded from four independent directions — see the OWASP crosswalk and owasp-mapping.

The decision path

The taxonomy navigator is organised as six questions about the system under review rather than as a flat list (source: Agentic-AI-Threats-and-Mitigations-1.1.pdf, pp. 21–32):

StepQuestionThreats
1Does the agent independently determine its own steps?T6, T7, T8
2Does it rely on stored memory for decisions?T1, T5
3Does it execute actions via tools, commands or integrations?T2, T3, T4, T11, T16, T17
4Does it rely on authentication of users, tools or services?T9
5Does it require human engagement to function?T10, T15
6Does it rely on multiple interacting agents?T12, T13, T14

Answering “no” at a step prunes its threats. This is what makes the document usable as a checklist rather than a reading list — a single-agent tool-using copilot with no persistent memory is scoped to steps 1, 3, 4 and 5.

Two internal inconsistencies

Flagged for anyone citing this document precisely:

  1. The introduction promises “threat models for four example scenarios” (p. 3) but only three are delivered: Enterprise Co-Pilots, Agentic IoT in Smart Home Security Cameras, and Agent-driven RPA (pp. 43–49).
  2. Playbook scope disagrees with the playbook mapping table. The overview table (p. 34) assigns Insecure Inter-Agent Protocol Abuse to both Playbook 3 and Playbook 4; the playbook headers themselves (pp. 37, 39) omit it from both. Playbook 6 is also labelled “Step 5: Multi-Agent System Threats” when the navigator places multi-agent threats at Step 6.

Neither affects the substance, but both mean the document is not yet self-consistent enough to auto-generate a control matrix from.