The Wire · Showcase
Two security patches land across OpenAI SDKs
By RepoJournal · Filed · About OpenAI
OpenAI's SDKs shipped two security-critical patches overnight, and neither can wait.
Both openai-node and codex closed holes that could leak credentials or let canceled tool calls keep running. openai-node redacts invalid Bedrock bearer credentials from HTTP-header errors and logs [1], while codex enforces subagent ownership across app-server inputs so parent-owned agents can't be mutated directly [2]. The tool-callback fix [3] is the one to drop everything for: it stops privileged parsed callbacks from firing after cancellation, closing a real exploit window in runTools(). Meanwhile codex honors response budgets when reading skill resources [4], hardening project config during credential brokering [5], and respecting disabled managed network config [6]. On the compatibility front, openai-python locks in support for large HTTP payloads with a 32 MiB+ contract test [8], so security fixes can't silently break big responses. The Node team also tightened Zod structured outputs to a fail-closed JSON-native subset [7], rejecting Date, BigInt, and other non-JSON types.
One email a day. Unsubscribe in one click.
Keep up with OpenAI in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
One email a day. Unsubscribe in one click. Read a past issue →
Action items
- → Upgrade openai-node to the latest patched release before your next deploy openai/openai-node [immediate]
- → Review and update your codex usage to ensure subagent ownership is enforced openai/codex [immediate]
- → Verify your openai-python version passes the large payload contract test openai/openai-python [plan]
References
- [1] fix(security): redact invalid Bedrock bearer credentials (#2419) openai/openai-node
- [2] Enforce subagent ownership across app-server inputs (#40464) openai/codex
- [3] fix(security): stop delayed and additional tool callbacks after cancellation (#2430) openai/openai-node
- [4] Honor response budgets when reading skill resources ↗ openai/codex
- [5] Harden project config when credential brokering is active ↗ openai/codex
- [6] Honor disabled managed network configuration ↗ openai/codex
- [7] fix(zod): support only provably JSON-native structured outputs ↗ openai/openai-node
- [8] test: protect large HTTP payload compatibility (#3721) openai/openai-python