Security Policy

Supported versions

Evarness is pre-1.0. Only the latest release (and main) receive security fixes.

Version	Supported
latest release / main	✅
older releases	❌

Reporting a vulnerability

Please do not open a public issue for security problems.

Report vulnerabilities privately via GitHub Security Advisories ("Report a vulnerability" on the repository's Security tab).

Include what you can of: affected component, a reproduction (graph + fixture + seed where applicable), and impact. You can expect an acknowledgment within a few days; fixes are coordinated before public disclosure.

Scope

Evarness currently executes harness graphs in simulation only: every model and tool response comes from a scripted fixture, and nothing in this package makes network calls or touches files outside its own data directory (~/.evarness/). That keeps the attack surface deliberately small, and it means:

• Simulation results are not a safety proof. A proof bundle demonstrates that declared invariants held over scripted scenarios, reproducibly. It does not establish that an agent is universally safe. Every bundle carries a "not proven" section stating exactly this.
• Signing proves integrity, not execution. A signed bundle proves the bundle is unaltered since signing — not that the runs happened as described. This is stated in the bundle and in verify output.

Reports that demonstrate a violation of a stated guarantee are very much in scope — for example: two runs with the same graph, fixture, and seed producing different canonical digests; an invariant verdict that alters a run's digest; a policy block that fails to prevent a simulated model call; a verify pass on a tampered bundle; or a trace event that silently drops a failure.