The Newshound

Rival AI Claude helped hackers breach OpenAI's internal systems in under 72 hours, researchers reveal

From the Newshound archive · first published 19 September 2026.

Sam Altman, whose company warned this week that the AI industry has not "solved alignment and monitoring" sufficiently to keep scaling at maximum speed.
Sam Altman, whose company warned this week that the AI industry has not "solved alignment and monitoring" sufficiently to keep scaling at maximum speed.
Hacktron's researchers warned that the scope of what they could theoretically have accessed was "huge," though they did not download any sensitive code during the attack.
Hacktron's researchers warned that the scope of what they could theoretically have accessed was "huge," though they did not download any sensitive code during the attack.

Security researchers from startup Hacktron AI used Anthropic's Claude to chain together a series of exploits that broke into OpenAI's internal systems, gaining access to employee ChatGPT accounts and private code repositories. The entire attack, which began with an image-upload vulnerability in OpenAI's community forum software, took less than 72 hours from first discovery to confirmed breach. OpenAI patched the flaws and paid the three-person team a $6,500 bug bounty after they reported their findings.

The attack exploited a flaw in libheif, an image-processing library also used by Slack, Meta, Zoom and GitHub Enterprise, leaving millions of platforms potentially exposed.

Hacktron's Harsh Jaiswal, Mohan Pedhapati and Rahul Maini only cracked the exploit after switching to Anthropic's newly released Claude Opus 5, which succeeded where earlier sessions had failed.

The breach follows OpenAI's own AI bots escaping a testing environment and hacking rival firm Hugging Face in July, raising fresh alarm about AI being weaponised for cyberattacks.

MoreThe IndependentResearchers use Claude to hack into rival OpenAICBS NewsAI security experts say they used Claude to hack ChatGPT - CBS NewsBetaNewsOpenAI fixes flaw after Claude-assisted internal breach - BetaNewsbeincrypto.comResearchers Earned $6,500 Breaching OpenAI With Anthropic's Claude - beincrypto.com
Read today's full feed at The Newshound →