Murat BekgiPenetration tester
Subscribe
Write-ups

Week 40 in tech: The agents' paper trail, 500,000 faces for zero arrests, and F-Droid 2.0

How OpenAI's agents left a million public links while hacking Hugging Face, London's facial recognition trial results, and F-Droid's big release.

Published
Reading time
1 min
Topic
Weekly

September 23 to 29. AI agents doing things nobody asked them to, and a surveillance system doing very little of what it was bought for.

The agents’ paper trail. Researchers rebuilt more than 80,000 attack payloads from July’s Hugging Face intrusion by OpenAI’s research agents, using nearly a million public short links the agents created and nobody cleaned up (Unite.AI, SwarmTraces). The deep dive is about the lesson underneath: a sandbox that can make GET requests to anywhere is not read-only.

500,000 faces, zero arrests. Freedom of information documents show the British Transport Police scanned more than 500,000 faces at London railway stations between February and July with live facial recognition. The system raised one alert, and it was a false match. No arrests came from the system, the trial cost more than £320,000, and the force is expanding it to Underground stations anyway (Biometric Update, The Guardian). Worth keeping in mind next time a vendor quotes you an accuracy rate: ask what the denominator is and what happened after each alert.

Worth a read: Two-tier encryption in the UK.

Ad tech, now with AI. The EFF says DraftKings is using AI to behaviorally target chronic gamblers.

In court: a US appeals court upheld the Pentagon’s designation of Anthropic as a supply chain risk.

Outside security: F-Droid 2.0 is out, the Dutch government is building an alternative to Microsoft based on NixOS, and Italy’s parliament voted to return to nuclear power.


New write-ups and both weekly posts land in the inbox — subscribe, or grab the RSS feed.