AgentCodec runs LLM workloads and ships an anonymous-telemetry path. Both surfaces have failure modes that affect users' data, money, and trust. We treat them seriously and we want to hear from you.

Please do not file public issues for security or privacy problems.

The preferred channel is GitHub private vulnerability reporting: on the public repo, open the Security tab and click "Report a vulnerability". This opens a private advisory only you and the maintainers can see, lets us collaborate (and test a fix on a private fork) without tipping off attackers, and handles CVE issuance and reporter credit on disclosure.

If you don't have a GitHub account or prefer email, write to [email protected] instead. Either way, include:

• A description of the issue and the impact you observe (or expect).
• A minimal reproducer if you can — Python snippet + YAML config.
• The AgentCodec version (python -c "import agentcodec; print(agentcodec.__version__)").
• Whether the issue is in the published library, the SemKNN backend, or the telemetry collector.

We aim to acknowledge within two business days and to ship a fix or a written plan within fourteen calendar days of confirming the report. We will credit the reporter in the changelog and the release notes unless asked otherwise.

• Anything that leaks prompt text, model output, reference answers, API keys, or user identifiers through:
  • The remote SemKNN call payload (/route body).
  • The anonymous telemetry payload (/telemetry body).
  • The ReliabilityResult trace (e.g. an accidental field with raw text under an "innocent" key name).
• Anything that causes the library to make HTTP calls the user did not authorize (e.g. an outbound request to an unexpected host).
• Anything that disables AGENTCODEC_TELEMETRY=0 or telemetry.enabled: false — those are contractual kill switches.
• Anything that bypasses the schema-pin denylist in agentcodec.telemetry._scrub() (the privacy fuse).
• Cost-accounting bugs that lead to wildly wrong cost_usd without the default_fallback warning firing — financial-impact bugs.
• Code execution / deserialization issues in the YAML / JSON loaders.

• Reports about the underlying LLM provider's API behaving badly (rate limits, content policy, hallucinations). Report those upstream.
• Vulnerabilities in pip install-ed dependencies — those go to the upstream project.
• DoS via "I sent a 10 MB prompt and the request was slow."

We document these so they're not surprises:

• Cancellation does not abort in-flight HTTP requests. The library uses worker threads for sync LLM SDK calls; cancelling the consumer drops the response but does not cancel server-side work. You may be billed for tokens generated after cancel.
• Sentence embeddings are lossy but not provably zero-leakage. The prompt-embedding the SemKNN router sends to the backend is a unit-norm BGE vector. Research-grade inversion attacks (vec2text) can sometimes recover topic-level content. The bge-small default reduces this. If your environment requires zero embedding leakage, set AGENTCODEC_TELEMETRY=0 and avoid the SemKNN router (use fixed or acm_table).
• The default_fallback cost tier is a stub. If a model's pricing isn't in MODEL_COSTS and the user didn't set cost_per_1m, the library uses $2/$8 and logs a loud warning. This is intentional but documented as a gotcha.

For security purposes, only the latest released version is supported. There are no LTS branches at this stage. If you're using an older version, the upgrade path is usually trivial; if not, file an issue and we'll help.

GitHub private vulnerability reporting is already a confidential channel, so no extra encryption is needed when you use it.

For email reports, [email protected] accepts PGP if you prefer. We will publish the public key on the repo and the website before the first public push; until then, please send a short email asking for it.