OWASP Agentic Top 10 coverage
Spec27 provides red-team coverage for risks described in the OWASP Top 10 for Agentic Applications 2026. Our Red Team evaluations help identify policy violations, examine the conversations that reveal them, and assess improvements after changes to an agent's model, prompts, or safeguards.
OWASP risk coverage
Spec27 does not aim to cover all ten categories. The table below outlines coverage for specific scenarios within each category, including any required integration. Follow the testing guide to configure and run the supported tests.
- Supported Spec27 supports testing the described scenarios through specifications, red-team tests, and response evaluation.
- With integration Testing the described scenarios in Spec27 uses specifications and evaluations together with target-system setup and evidence, such as test identities, tool traces, or audit records.
- Not covered The described scenarios fall outside Spec27's current testing scope.
| OWASP risk | What the risk means | Spec27 coverage |
|---|---|---|
| ASI01: Agent Goal Hijack Supported | Attacker-controlled input redirects the agent's objective, decisions, or subsequent behavior. | Spec27 tests whether adversarial instructions redirect the conversation or elicit policy-breaking replies. Indirect injection tests require a controlled document or website that the target retrieves. How to test goal hijacking? |
| ASI02: Tool Misuse and Exploitation With integration | An agent applies a legitimate tool in an unsafe or unintended way within its existing authority. | Spec27 exercises tool-misuse scenarios and verifies expected calls through telemetry. Checking prohibited calls and completed actions requires target-system audit evidence. How to test tool misuse? |
| ASI03: Identity and Privilege Abuse With integration | An agent misuses identities, inherited credentials, or delegation to cross an access boundary. | Spec27 tests responses for unauthorized disclosure using customer-configured identities and protected synthetic data. Verifying actions requires target-system audit evidence. How to test access boundaries? |
| ASI04: Agentic Supply Chain Vulnerabilities Not covered | Compromised tools, plugins, models, templates, registries, or other dependencies enter the agent's execution chain. | Dependency integrity, registry compromise, and update-channel security require a supply-chain assessment. Spec27 can test responses to injected content through the ASI01 scenario. |
| ASI05: Unexpected Code Execution (RCE) Not covered | Untrusted input causes unintended code execution, potentially compromising a host or sandbox. | Confirming execution or sandbox compromise requires an isolated execution environment and process or filesystem evidence. Code-generation behavior can be evaluated separately as a response-policy test. |
| ASI06: Memory & Context Poisoning Supported | Malicious information retained or reused by the agent changes later reasoning or behavior. | Spec27 tests whether a planted instruction affects subsequent replies within a conversation. Persistent and cross-user memory tests additionally require customer-managed memory, session setup, and retrieval evidence. How to test context poisoning? |
| ASI07: Insecure Inter-Agent Communication Not covered | Agent messages are spoofed, modified, replayed, or exposed across inadequately protected communication channels. | Authentication, integrity, and replay protection need a real inter-agent communication test environment. |
| ASI08: Cascading Failures Not covered | An initial fault propagates and amplifies across agents, tasks, sessions, or workflows. | Assessing propagation and downstream impact requires a system-level fault-injection and tracing environment. |
| ASI09: Human-Agent Trust Exploitation Supported | An agent's apparent authority or persuasive output induces people to disclose information or approve unsafe actions. | Spec27 evaluates replies for credential requests, fabricated approval, or pressure to take unsafe actions. This covers the agent's advice; actual human decisions require a separate assessment. How to test manipulative advice? |
| ASI10: Rogue Agents Not covered | An agent persistently deviates from its authorized purpose, for example through deception, reward hacking, or harmful coordination. | Autonomous drift, coordination, and containment require sustained execution and governance tests. |
* Risk descriptions adapted from the OWASP Top 10 for Agentic Applications 2026 (CC BY-SA 4.0).