XNews

Researchers Hack OpenAI Using Claude Opus 5, Take Over Employee Accounts

Researchers Hack OpenAI Using Claude Opus 5, Take Over Employee Accounts

Three Hacktron researchers exploited a libheif heap overflow in OpenAI's community forum and an SSO flaw to take over multiple OpenAI employee ChatGPT accounts. They used one employee's Codex to open a harmless PR in the internal repo, spending under $3,000 in tokens. OpenAI fixed the issues and paid a $6,500 bounty.

Key insights

Attack chain began with a malicious image

The attack began with a malicious HEIC image posted to OpenAI's official community forum, which runs on the open-source Discourse software. Discourse sends HEIC files to ImageMagick and libheif for conversion. libheif had a heap overflow patched upstream, but the fix was not marked as a security update and received no CVE, so Debian 12 and 13 did not pull the patch. Because Discourse's Docker image is based on Debian 12, uploading the crafted image gave the researchers arbitrary code execution on the forum server.

SSO misconfiguration led to account takeover

After taking the forum server, the researchers exploited an OpenAI single sign-on misconfiguration. The forum supported 'Sign in with OpenAI account', so active forum users' ChatGPT and Codex accounts could be taken over without user interaction. One compromised employee had Codex connected to OpenAI's GitHub organization. To demonstrate access without touching sensitive data, the researchers had that Codex open a harmless pull request in OpenAI's internal openai/openai monorepo, then stopped. Hacktron said any service connected to OpenAI SSO would have the same exposure, including GitHub, Slack, and email.

Claude Opus 5 completed the exploit quickly

Hacktron's three researchers started with Opus 4.8, which found the unpatched libheif code but could not produce a working exploit under ASLR. After Anthropic released Opus 5, a fresh session completed the exploit in about three hours on a local Mac, and it worked against a test Discourse instance the next day. Claude generated most of the attack code. Opus initially refused to write an exploit for the real remote target, so the researchers disguised their own test instance as a CTF challenge. Total token spending for the HEIF Heist project, which also included work against Slack, Zoom, and Meta, was under $3,000.

Timeline, fix, and disclosure

The timeline from initial discovery to entering OpenAI's repository was under 72 hours. OpenAI confirmed a fix about 14 hours after receiving the report. Discourse received the report on a Saturday, shipped a patch with image-processing sandboxing on Monday, and published a security advisory on July 28. OpenAI paid $6,500 on September 1 and noted that the forum had been excluded from its bug bounty program; the payment covered only OpenAI-specific findings, not the Discourse vulnerabilities. Hacktron published full technical details on September 13, followed by Wall Street Journal coverage.

Broader exploitation campaign

Hacktron described the broader project as running for about two months and costing less than $3,000 in total token spend. Adapting an exploit to a new company usually takes one or two days, often without knowing the target's libheif version or deployment environment. The researchers sent thousands of images and repeatedly crashed image-processing services; to their knowledge, only Shopify detected the attempts. They also noted that switching from Opus 5 to OpenAI's GPT-5.6 Sol produced another clear capability gain in the unknown-target scenario.

Human guidance still played a central role

Hacktron said the operation was not fully autonomous; experienced human guidance still played an important role. The attack combined Claude-generated exploit code, manual reconnaissance, and step-by-step adaptation to the target. The researchers chose a harmless PR to prove reach and avoid data exposure, and they deliberately stopped after demonstrating access. The disclosure shows that frontier models can accelerate vulnerability chaining, but the process still depends on human operators deciding where to direct the agents.

Source posts