After spending billions, OpenAI still has gaps in its cybersecurity
Hacktron accessed OpenAI employee accounts via chained flaws, and Accomplish escaped the Codex sandbox; both were fixed.
CSO Online reports two coordinated disclosures of security weaknesses at OpenAI. On July 25, 2026, Hacktron researchers chained an image-processing remote-code-execution flaw with stolen authentication tokens to access multiple employees’ ChatGPT accounts and internal repositories; they performed only a benign action, did not exfiltrate data, and said the issues were fixed. They used Anthropic’s Claude iteratively for reconnaissance, analysis, and exploit development. Separately, Accomplish reported two Codex coding-agent sandbox escapes on August 12, 2026, both fixed within eight days.
- Hacktron chained an image-library RCE flaw to steal tokens on July 25, 2026.
- Employee ChatGPT accounts exposed internal OpenAI repositories; no data was stolen.
- Claude was used to refine reconnaissance and the exploit chain.
- Two Codex sandbox escapes reported August 12 were fixed within eight days.
From the original article
4 min read at csoonline.com
Two separate reports of security flaws in OpenAI systems highlight how even a company spending billions on developing its own AI-powered cybersecurity testing tools remains vulnerable. In one incident, researchers breached OpenAI systems with the help of a rival AI developer’s tools, while another group of researchers tricked OpenAI’s Codex agent into bypassing its sandbox controls.
csoonline.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.