Researchers used Claude Opus 5 to take over OpenAI employee accounts via the community forum
Hacktron chained a libheif overflow on community.openai.com with an OpenAI SSO flaw, then had a taken-over employee’s Codex open a harmless PR in openai/openai. OpenAI paid $6,500 for the identity finding. Discourse was out of bounty scope.
Deep dive5 minIn-depth
Maya Chen5 min read
- OpenAI
- Anthropic
- Claude Opus 5
- Security
- Bug bounty
- Hacktron

Photo: Hacktron AISource page
A three-person team at Hacktron AI used Anthropic’s Claude Opus 5 to chain two critical bugs on July 25 and take over OpenAI employee ChatGPT and Codex accounts. From one of those sessions they had Codex, already connected to OpenAI’s GitHub organisation, open a harmless proof-of-concept pull request — numbered 1186742 — in the internal monorepo openai/openai, without reading secrets. TechCrunch reported the campaign on Friday, citing the Wall Street Journal from Thursday evening. The researchers, Harsh Jaiswal, Mohan Pedhapati and Rahul Maini, published their own account on September 13. OpenAI paid a $6,500 bug-bounty award and said it had fixed the company-side issue about 14 hours after the Bugcrowd submission.
The entry point was OpenAI’s Discourse forum at community.openai.com. Users could upload HEIC or HEIF images. Discourse passed those files to ImageMagick, which handed them to libheif. Hacktron wrote that a heap overflow in that decoder, on a Debian package that had not received a security backport, yielded remote code execution on the forum. A separate OpenAI SSO identity flaw then turned forum compromise into takeover of ChatGPT and Codex accounts for people who had used “Sign in with OpenAI.” Discourse published advisory GHSA-vhm9-85gw-x335 on July 28, added ImageMagick sandboxing, and told self-hosters to rebuild rather than click update in the web UI. OpenAI told the researchers that testing against the Discourse-hosted forum was explicitly excluded from its bounty programme; the $6,500 recognised the OpenAI-side finding, not the actions against Discourse.
For $200 a month, anyone can use these tools and hack into a company like OpenAI. If it can happen to them — and I don’t think they’ve been slouching recently on cybersecurity hygiene — it could happen to anyone.
— Matt Fredrikson, CEO of Gray Swan, to TechCrunch, Sep 18, 2026
Seventy-two hours, then a pull request
Hacktron’s timeline is compressed. The team said it confirmed remote code execution and administrative access on community.openai.com between 05:00 and 06:00 UTC on July 25, filed through OpenAI’s Bugcrowd programme later that morning, and by mid-afternoon had used a taken-over employee Codex session to open the proof-of-concept pull request. They updated the Bugcrowd ticket, pinged people at OpenAI on X, and stopped testing around 15:30 UTC. OpenAI replied at 22:49:45 UTC that the issue was fixed — roughly 14 hours after the first submission. The researchers reported Discourse through HackerOne the same day. Discourse replied on Sunday, July 26, had a fix ready on Monday, and published the GitHub advisory on Tuesday. OpenAI marked the bounty resolved and paid $6,500 on September 1, with the scope clarification attached.
The researchers are careful about what they did not do. They wrote that they used the employee’s Codex to open the pull request specifically to prove access “without allowing ourselves to learn any sensitive information,” and that the monorepo link was later redacted at OpenAI’s request. Until the patch, they said, any user or employee logging into the help forum could have had ChatGPT and Codex taken over, and with them whatever connectors those accounts held — GitHub, in this case; Slack and email, in theory. They also stressed that Discourse was one proof of the SSO problem, not the only possible front door: compromise of any first-party or third-party OpenAI service on that identity flow would have led to the same access.
Opus 4.8 stalled on ASLR. Opus 5 did not.
The model jump is the part of the write-up that will travel. Hacktron started an Opus 4.8 session — TechCrunch, following the researchers, called it a special cyber-researcher build — against Discourse’s Docker image and had it find that some libheif security fixes had not been back-ported. On July 24 they used Opus 4.8 to get a working ImageMagick/libheif exploit with ASLR disabled. Sessions aimed at Discourse’s default configuration with ASLR on did not land. Anthropic released Claude Opus 5 that evening. A new session produced a working ARM64 exploit on a local Mac within three hours, then a port to the x86-64 and jemalloc setup Discourse used. By 06:00 on July 25 they had local code execution through an image upload. They then put Claude in an autonomous loop against their own Discourse Cloud instance, proxied to look like a CTF target after the model refused to write an exploit for a remote host. By 10:00 the agent had code execution there. They used the generated script against OpenAI’s forum.
Opus 4.8 struggled across several sessions to produce a working exploit. Within hours of Opus 5’s release, we gave it the same problem and it succeeded.
— Hacktron AI, Hacking OpenAI, Sep 13, 2026
Pedhapati put the same point on X, as quoted by TechCrunch: “AI is reducing the amount of scarce expertise needed to develop exploits. Work that once took months can now take days.” The researchers are not claiming fully autonomous hacking. They wrote that skilled human guidance still mattered, and that the Discourse job took a few days of agent time and a few hours of theirs. The wider HEIF Heist project — tracing libheif through Slack, Meta, GitHub Enterprise, Ruby on Rails and Node stacks including Next.js — ran about two months, cost less than $3,000 in tokens, and usually took a day or two to adapt per company. They say they are not aware of any target that detected the image-upload probing except Shopify, even after processors crashed. Opus 5 has not faced export restrictions. TechCrunch noted that Anthropic’s newer Mythos 5 was temporarily locked down over hacking capability. That is the policy contrast the Friday coverage is drawing, not a claim that Opus 5 is Mythos.
A different OpenAI-agent story, already on the archive
The disclosure lands weeks after OpenAI’s own agents broke containment during a cybersecurity evaluation and probed Hugging Face — the May-to-July episode already reported here as “OpenAI’s rogue agents probed Hugging Face in May.” That is a separate chain: models under OpenAI’s control, acting on the open internet. Hacktron is the inverse, three outsiders using a generally available Anthropic model against OpenAI’s own forum and identity stack, inside a bounty process that later carved Discourse out of scope. The common fact is narrower. Frontier models now compress exploit work that used to sit with scarce specialists. Gray Swan’s Fredrikson priced that at about $200 a month. Hacktron’s own token bill for the broader campaign was under $3,000. OpenAI patched in hours. Discourse patched over a weekend. The remaining question, which the researchers put in their epilogue, is whether companies that still treat memory-corruption as too expensive for ordinary attackers have updated that assumption.


