Agentic Evidence is an OWASP open-source reference project for collecting, correlating, and evaluating security evidence from agentic AI execution. It focuses on reconstructing actions and determining what the evidence supports an investigator concluding.
Agent logs and traces can show that events were recorded, but may not establish which identity authorized an action, what input influenced it, which tool descriptor was used, or whether relevant records are missing or untrustworthy. The project explores these evidence gaps through reproducible scenarios.
A conventional observability dashboard primarily helps operators understand system behavior and performance. Agentic Evidence adds a security-investigation perspective: evidence provenance, authorization context, cross-event correlation, expected-versus-observed evidence, and explicit limits on what can be concluded.
It builds on established telemetry where appropriate. It does not claim that telemetry alone proves an action was trustworthy.
No. The initial deliverable is a local reference implementation for learning, experimentation, and repeatable demonstrations. Production deployment, high availability, access controls for multi-user environments, retention controls, and integrations require separate design and validation.
Not by itself. The initial focus is evidence collection and investigation, not enforcement or automatic remediation. Authorization controls and runtime defenses must be implemented by the system that executes the agent.
The initial scenarios are intended to be deterministic and fixture-driven. A live model should not be required to reproduce the baseline, retrieval-poisoning, or tool-descriptor-modification scenarios. Future integrations may be considered after the evidence model and core tests are stable.
- Benign execution to establish a baseline.
- Retrieval poisoning to examine whether retrieved content and subsequent actions can be correlated.
- Tool-descriptor modification to examine descriptor identity, versioning, integrity, and provenance evidence.
These are controlled demonstrations, not claims of comprehensive attack coverage.
Evidence is a record or artifact used to support or challenge a claim about an execution. Examples include a tool invocation event, a policy decision, a trace span, a descriptor hash, or a provenance statement.
Evidence has different levels of trust. A component-generated log is an assertion from that component unless there is a separate mechanism to validate it.
No. A trace reflects what instrumented components emitted and what the collection pipeline retained. A compromised or faulty component may emit false events, omit events, or use incorrect identifiers. The project should expose those limitations instead of treating trace presence as proof.
No. A hash can help identify whether bytes differ from a known value. It does not prove who produced the descriptor, whether the reference hash is trustworthy, whether the descriptor was actually used, or whether relevant events were omitted. Stronger conclusions require trustworthy reference values and appropriate provenance or verification mechanisms.
An evidence gap is a missing or insufficient record needed to support a security conclusion. Examples include a tool invocation without a correlated authorization decision, a descriptor without a known version, or an execution timeline with missing events. A gap should be reported explicitly rather than silently filled with an assumption.
No. The intended role is complementary. Taxonomies describe risks and testing projects exercise security properties. Agentic Evidence focuses on the evidence available to reconstruct execution and evaluate investigation claims.
The proposed baseline is Python/FastAPI, versioned JSON and JSON Schema, OpenTelemetry, local SQLite or JSON fixtures, a lightweight investigation UI, Docker Compose, pytest, and GitHub Actions. These choices may change through documented architecture decisions as the implementation develops.
The initial local MVP is intended to run with Docker Compose and synthetic data. It should not require Kubernetes, a cloud account, or a live model API key.
Do not submit sensitive production traces, credentials, secrets, customer data, or personal data to public issues or pull requests. Use synthetic fixtures or properly sanitized data. Follow your organization's data-handling requirements.
Start by checking the milestones, the contribution guidelines, and open issues. Useful contributions include evidence-schema design, deterministic scenarios, telemetry correlation, UI improvements, test cases, documentation, and independent validation. Discuss substantial changes in an issue before opening a pull request.
Use the private reporting method documented in SECURITY.md. Do not disclose an unpatched vulnerability in a public issue.
The project is intended to use Apache License 2.0. The repository's
published LICENSE file is authoritative. Check it before
redistributing code.
Use the official OWASP project listing and repository milestones, issues, and release notes. This documentation describes the intended MVP and should not be treated as proof that a planned feature has already shipped.