Tripwire is pre-1.0 and released from main. Security fixes land on the latest 0.1.x release on PyPI (tripwire-eval); there is no backporting to older versions.

Please report suspected vulnerabilities privately, not in a public issue.

• Preferred: open a private report via GitHub Security Advisories on this repository (the "Report a vulnerability" button under the Security tab).
• Alternatively: email the maintainer at arunjay101@gmail.com.

Include enough detail to reproduce: affected version, environment, a minimal example, and the impact you observed. You can expect an initial acknowledgement within about a week. Once a fix is available, the advisory will be published and the reporter credited unless you prefer to stay anonymous.

Tripwire is a red-team evaluation harness: it deliberately generates and runs prompt-injection attack payloads against the agent under test. Those payloads, the adaptive attackers, and the fact that a poorly defended agent leaks its canary are the intended behavior of the tool, not vulnerabilities.

The bring-your-own-agent path (tripwire --agent path/to/agent.py) imports and executes the Python file you point it at. That is by design and equivalent to running your own script; only pass files you trust.

In scope for a report are security issues in the harness itself, for example:

• an unintended way for untrusted input (a config file, a scenario, a tool result, an attacker payload) to execute code or read/write files outside the harness's normal operation,
• a path-traversal or injection flaw in how Tripwire loads configs, agents, or results,
• a vulnerability introduced by one of Tripwire's own dependencies as used here.

When in doubt, report it privately and we will triage.