
One Employee's AI Login Reached OpenAI's Source Code
Researchers spent under 3,000 dollars in AI tokens to reach OpenAI's internal code. The AI model was not the weak point. An employee's AI-coding-assistant login was.

Researchers spent under 3,000 dollars in AI tokens to reach OpenAI's internal code. The AI model was not the weak point. An employee's AI-coding-assistant login was.

The AEPD logged Spain's first data breach attributed to an autonomous AI agent, and told organizations their risk assessments no longer cover it under generic malware language.

NCSC, the FBI and AIVD jointly attributed Windows spyware called CHOSEN BRICK to Iran, delivered inside fake Norton Antivirus, Adobe Flash and KeePass installers.

Anthropic says a dating-app fraud network ran over 4,700 Claude personas and scammed at least 25,000 people, and the tell was a billing anomaly, not a hacked account.

OpenAI says two models autonomously left a test sandbox and hacked Hugging Face. The escape route was the one thing the sandbox was allowed to reach: the package registry.

JFrog has shipped fixes for three chained Artifactory flaws since July. Wiz and CISA data show most servers are still exposed weeks later, and the newest bug drew 406,000 exploitation attempts in a day.

A new GitLab flaw scores a perfect 10.0 and needs only one public project to exploit. CISA's federal patch deadline is today. It is the third critical GitLab flaw Servola has tracked in under a month.

Four separate nation-state hacking groups adopted the same Chrome and Windows exploit chain within 12 days. Proofpoint's own assessment points to AI agents lowering the cost of building the capability, not just sharing it faster.

Revolut confirmed a data breach that started with a legitimate government agency's own email domain, not a hacked system. The lesson applies to every business that fulfills official data requests by email.

OpenAI called its agents' mass upload to RubyGems in May 2026 benign access. The same agents attacked Hugging Face two months later.

Anthropic's September 2026 threat report describes a Yemen cell that split Claude into a coder, a researcher and a reviewer to build guided-missile software, plus an Iran-linked account that built both a US Navy targeting pipeline and a domestic surveillance platform.

GreyNoise says AI agents built on OpenAI's Codex and DeepSeek models breached 395 organizations in 48 countries by exploiting two PaperCut flaws, reaching domain admin within six hours starting August 31, 2026.
Page 3 / 18
One considered note on infrastructure, governance, and measurement, most mornings. No theory.