DevelopingSecurity

Hacktron researchers used Anthropic's Claude to reach OpenAI accounts and internal repository in 72 hours

CryptoSlate reported that Hacktron researchers used Anthropic's Claude to chain two flaws, reach OpenAI employees' ChatGPT and Codex accounts and open a pull request in the openai/openai monorepo within 72 hours.

Why it matters

CryptoSlate's account points at permission design rather than a single bug: coding agents connected to corporate accounts concentrate access, so one compromised employee account exposed GitHub and internal repositories. OpenAI reportedly patched the identity flaw within about 14 hours and paid a bounty, so the reported exposure was short-lived, and the significance beyond that single outlet's reporting is limited.

Key facts

Researchers at the cybersecurity startup Hacktron chained an image-processing vulnerability with a flaw in OpenAI's identity infrastructure in July to reach multiple employees' ChatGPT and Codex accounts.

Reported by CryptoSlate

Three Hacktron researchers used Anthropic's Claude to reach OpenAI accounts and an internal code repository within 72 hours.

72 hoursReported by CryptoSlate

One compromised Codex account was connected to OpenAI's GitHub organization, which gave the researchers a path into the company's internal software environment.

Reported by CryptoSlate

The team stopped after instructing the compromised employee's Codex account to create a harmless pull request inside OpenAI's private openai/openai monorepo.

Reported by CryptoSlate

Hacktron said the researchers did not inspect proprietary source code.

Reported by CryptoSlate

Hacktron disclosed the vulnerabilities and ended further testing in the week before CryptoSlate's report.

Reported by CryptoSlate

OpenAI reportedly fixed the identity-side flaw roughly 14 hours after receiving the report and later paid Hacktron a bounty.

6,500 USDReported by CryptoSlate

CryptoSlate wrote that the attack accelerated after Anthropic released Claude Opus 5, which cleared an exploitation hurdle its predecessor had repeatedly failed to clear.

Reported by CryptoSlate

What happened

CryptoSlate reported that three researchers at the cybersecurity startup Hacktron used Anthropic's Claude to reach multiple OpenAI employees' ChatGPT and Codex accounts and an internal code repository within 72 hours. The researchers chained a flaw in an image-processing system with a separate flaw in OpenAI's identity infrastructure in July, and one compromised Codex account was connected to OpenAI's GitHub organization, which gave the team a path into the company's internal software environment, CryptoSlate wrote. The team stopped after directing the compromised employee's Codex account to create a harmless pull request in OpenAI's private openai/openai monorepo, and Hacktron said the researchers did not inspect proprietary source code. Hacktron disclosed the vulnerabilities and ended further testing in the week before CryptoSlate's report. CryptoSlate said OpenAI reportedly fixed the identity-side flaw about 14 hours after receiving the report and later paid Hacktron a $6,500 bounty, and that the attack accelerated after Anthropic released Claude Opus 5, which cleared an exploitation hurdle its predecessor had repeatedly failed to. Only CryptoSlate has reported this, and the account of OpenAI's response and of Opus 5's role rests on that outlet alone.

How the story developed

  1. First report

    First report by CryptoSlate

    Anthropic’s Claude helped 3 researchers breach OpenAI in under 72 hours

    CryptoSlate

  2. Update

    Image-processing flaw chained with an OpenAI identity flaw

    Hacktron researchers combined the two vulnerabilities to reach multiple employees' ChatGPT and Codex accounts, which CryptoSlate placed in July.

    CryptoSlate

  3. Resolution

    OpenAI reportedly patched the identity flaw in about 14 hours and paid a bounty

    CryptoSlate reported the fix and the $6,500 payment to Hacktron; OpenAI has not been reported confirming either.

    CryptoSlate

  4. Update

    Anthropic released Claude Opus 5

    CryptoSlate linked the release to faster exploitation, saying Opus 5 cleared a hurdle its predecessor had repeatedly failed to clear.

    CryptoSlate

  5. Update

    Researchers stop after a harmless pull request in openai/openai

    The compromised employee's Codex account was asked to open a harmless pull request in OpenAI's private monorepo, and the team halted there.

    CryptoSlate

  6. Update

    CryptoSlate publishes the account

    The outlet is the only source in the desk's set for this story.

    CryptoSlate

Affected entities

Original sources

PublisherReportRolePublished
CryptoSlate
Crypto media
Anthropic’s Claude helped 3 researchers breach OpenAI in under 72 hoursOriginal
Hacktron researchers used Anthropic's Claude to reach OpenAI accounts and internal repository in 72 hours — Crypto News Intelligence