gpt2agent proxies a full ChatGPT Plus/Pro account, so security reports are taken seriously. Please read this before filing.
Do not open a public issue for a vulnerability. Instead use GitHub's private reporting: Security → Report a vulnerability on https://github.com/robotlearning123/gpt2agent/security/advisories/new.
Include: affected version/commit, a minimal reproduction, the impact, and any suggested fix. We aim to acknowledge within a few days.
gpt2agent talks to ChatGPT's private backend by impersonating the web client
(TLS fingerprint impersonation via curl_cffi, vendored Proof-of-Work + Cloudflare
Turnstile solvers). Some properties are inherent to that design and are documented,
not bugs:
- ToS / account risk. Using this likely violates the OpenAI Terms of Service; automated traffic can get an account rate-limited, challenged, or banned. Use an account you can afford to lose.
- Unauthenticated HTTP transport. The HTTP transport has no auth and exposes
the whole account. It binds
127.0.0.1by default and refuses non-loopback binds unlessGPT2AGENT_ALLOW_REMOTE=1is set. Prefer the stdio transport. - Limited PII redaction. Returned conversation/memory text strips only emails and phone numbers; treat output as non-anonymized.
In-scope reports we want to hear about: token/secret leakage in logs or errors,
ways to bypass the loopback/GPT2AGENT_ALLOW_REMOTE bind guard, injection that
makes a tool act on attacker-controlled data, or unsafe file/permission handling.
Only the latest released version is supported. Please reproduce against main
or the newest release before reporting.