OpenAI NewsSep 12, 2025, 20:00 UTC+8

Working with US CAISI and UK AISI to build more secure AI systems

OpenAI said its work with US CAISI and UK AISI found and fixed 2 novel ChatGPT Agent vulnerabilities; CAISI built a proof-of-concept exploit chain with about a 50% success rate, and OpenAI fixed it within 1 business day. The post says the bugs let attackers bypass protections under certain conditions, remotely control session-accessible systems, and impersonate logged-in users; UK AISI has red-teamed bio-misuse safeguards for ChatGPT Agent and GPT-5 since May 2025, but the truncated post does not disclose further results.

Why it matters: This is not generic safety PR. OpenAI discloses 2 new ChatGPT Agent vulns, ~50% CAISI PoC success, and a 1-business-day fix, so HKR-H/K/R all pass. Kept below 85 because the UK AISI section is truncated and the broader impact is not disclosed.

Read the original ↗Export Markdown

Take

The sharp number here is the ~50% exploit success rate, not the US-UK partnership wrapper. CAISI chained traditional software bugs with an AI agent hijacking attack, bypassed protections, remotely controlled session-accessible systems, and impersonated users on logged-in sites. That is no longer the toy version of prompt injection; it is account-bound agent security under real adversarial pressure.

OpenAI’s 1-business-day patch is good incident response, but it also shows the product problem: once ChatGPT Agent touches logged-in web state, model policy, sandboxing, and browser permissions collapse into one attack surface. UK AISI has red-teamed bio-misuse safeguards for ChatGPT Agent and GPT-5 since May 2025, but the truncated post gives no results. The disclosure gives one hard security win and leaves the scarier evals offstage.

Source: OpenAI News · openai.com
Published: Sep 12, 2025, 20:00 UTC+8 · Sep 12, 2025