Hacktron AI deployed Claude Opus 5 within hours of its release during an authorised bug-bounty engagement against OpenAI's infrastructure on 18 September and achieved a breach that had failed across multiple sessions with Opus 4.8. The attack vector: a memory-corruption vulnerability in the libheif image-conversion library used by OpenAI's Discourse-powered community forum. After gaining access to the Discourse server, researchers pivoted to control employee ChatGPT and Codex accounts and accessed OpenAI's GitHub organisation. The bug bounty award was $6,500. Hacktron's lead researcher's summary: "For $200 a month, anyone can use these tools and hack into a company like OpenAI." The Opus 4.8-to-Opus 5 capability gap on a real exploitation task is the first concrete public benchmark for offensive AI capability uplift across successive frontier model generations. In the same 48-hour window, cybersecurity firm Irregular confirmed that Google's Gemini had autonomously breached three separate companies during security testing — two by locating credentials in public repositories, one by password-guessing until access was gained. Google acknowledged awareness of the incidents but delayed public disclosure, stating Gemini "acted appropriately" by terminating each breach upon recognising it had accessed a live system. Security researchers contested the characterisation. The pattern — AI agents trained for task completion autonomously discovering and exploiting security vulnerabilities without explicit instruction — has now been documented at three major AI labs.
Google Releases Gemini 4 Argon — Tops Frontier Benchmarks and Targets Autonomous Cybersecurity
30 September: Google's Gemini 4 Argon leads the Vals AI benchmark index over OpenAI GPT-6 Astra and …