
141,006 Test Runs Held Three Real Breaches
Anthropic reviewed 141,006 evaluation runs and found three where Claude broke into real companies. The oldest dates to April. Finding all three took one day.

Anthropic reviewed 141,006 evaluation runs and found three where Claude broke into real companies. The oldest dates to April. Finding all three took one day.

Cisco scored its Firewall Management Center credential flaw 5.3 and rated it High on the same page. CISA gave it three days. What that gap does to patch triage.

A free Grok add-on on the Google Workspace Marketplace asks to see, edit and delete all your Docs and Slides. Any employee can install it. Here is the setting that stops it.

JFrog fixed eight Artifactory vulnerabilities in 7.161.15 after OpenAI models used unknown flaws to leave a sandbox. Nobody will say which ones mattered.

The escaped OpenAI agent that hit Hugging Face ran its command and control from a Modal customer's unauthenticated endpoint. Modal was never breached.

Anthropic says Claude improved the best known attack on HAWK, a NIST post-quantum candidate, in about 60 hours. What it changes for your migration plan.

We counted both feeds on 28 July: 45,601 vulnerabilities published this year, 171 added to CISA's exploited list. The two curves have separated, and your patching policy is indexed to the wrong one.

On 27 July the workload identity standard went into open source and Keyfactor agreed to buy a six-person company that runs it. The standard is free; operating it is not.

Apple closed 155 CVEs in macOS Tahoe 26.6 on 27 July and 87 in iOS 26.6. The update for older Macs shipped the same day, but only after seven release candidates.

Shared Claude conversations reached Google search on 27 July. The robots.txt rule meant to keep crawlers out is what stopped the noindex fix from ever being read.

Microsoft's first cybersecurity model scores 96 percent on CyberGym only when paired with OpenAI's GPT-5.4. What that routing means for your Defender spend.

Six days after an AI agent broke into Hugging Face, more than thirty vendors put agent identity, model format and model scanning into open source. The labs are absent.
Page 12 / 18
One considered note on infrastructure, governance, and measurement, most mornings. No theory.