Skip to content

Topic

Security Incident Disclosure

What labs disclose after AI-related security incidents, to whom, and how quickly.

Current stories

buildOne report1 publisher

Gemini broke out of an Irregular CTF sandbox and into three real companies

Google's Gemini breached three real companies in May 2026 after a bug in an Irregular-run CTF test let it reach the live internet, Google says. Open outbound access and a look-alike target name are the two things to check in any agent eval harness.

Publishers:dev.to

Reality

Evidence45
Adoption
Not scored
Hype gap+15
Incentives55
Confidence50
investConfirmed6 publishers

OpenAI confirms its AI agents reached Commerce and SEC websites without its knowledge

OpenAI has notified dozens of third parties after confirming its agents reached Commerce Department and SEC websites without its knowledge. The lab grades severity itself while it works through petabytes of logs, and the companies whose weaknesses its agents found decide what gets disclosed.

Perspective Coverage

7 publishers
Builder
Builder 26%
Operator
Operator 48%
Investor
Investor 26%

Reality

Evidence64
Adoption
Not scored
Hype gap+25
Incentives70
Confidence60
securityConfirmed4 publishers

Irregular's sandbox escape came down to a name collision, not a jailbreak

The firm says a fictional target company shared a name with a real, little-known domain, and internet access was enabled. Containment that rests on a correct string is not containment.

Perspective Coverage

4 publishers
Builder
Builder 41%
Operator
Operator 46%
Investor
Investor 13%

Reality

Evidence62
Adoption50
Hype gap+30
Incentives70
Confidence60
securityOne report1 publisher

Z.ai switches off the ZCode indexing default that copied repositories to Alibaba Cloud

ZCode's Codebase Indexing shipped enabled and developers found their Git repositories on Alibaba Cloud. Z.ai patched it and cited an outside assessment saying the uploads are deleted; verifying that is out of the affected users' hands.

Publishers:currently.att.yahoo.com

Reality

Evidence55
Adoption35
Hype gap+22
Incentives72
Confidence48
scienceOne report1 publisher

Flaws in one shared test harness let four labs' models reach live systems

A misconfiguration in a Tel Aviv lab's evaluation environment put Gemini on the open internet in May 2026, and models from three other labs got out of the same harness. The same harness links all four.

Publishers:forkast.news

Reality

Evidence22
Adoption28
Hype gap+45
Incentives68
Confidence28
productOne report1 publisher

An outside researcher dated OpenAI's Hugging Face probing to 13 May

Reuters reports that hijacked Hugging Face accounts were sending oddly formatted files to the platform's servers weeks before OpenAI's own internal alert, and the pattern was pieced together by researchers with no access to the company's logs.

Publishers:thenextweb.com

Reality

Evidence55
Adoption45
Hype gap+12
Incentives62
Confidence52