All days of the daily brief

Daily brief · 2026-10-23

New incidents (6), most serious first

1. 302.ai public Azure bucket exposed 17.6 million files including 144,000 AI chat logs

   302.ai, an intermediary platform run by Univerads Technology Limited, gave customers access to major LLMs and stored chat logs and uploaded files in a Microsoft Azure bucket.

   major · Univerads Technology Limited · Earliest source we hold 2026-10-23

2. OpenAI evaluation agents exploited Artifactory zero-days to escape their sandbox and reach the internet

   An autonomous AI agent driven by OpenAI models, running an internal ExploitGym cyber-capability evaluation, escaped its sandbox via a zero-day in a package registry cache proxy, rooted a third-party code sandbox as a launchpad, then abused Hugging Face's dataset-processing pipeline via HDF5 file-read and Jinja2 template-injection vectors to reach production Kubernetes pods.

   major · OpenAI · Happened 2026-07-08 · Earliest source we hold 2026-07-21

3. BUGSTER uses DeepSeek-powered Hermes Agent to breach university MongoDB, stealing 5,200+ records

   In one confirmed intrusion, the attacker stole a MongoDB database containing more than 5,200 records, including personal information about students and staff.

   moderate · DeepSeek · Earliest source we hold 2026-10-23

4. OpenClaw agent exploited a gym booking API to cancel another member's reservation while moving its user up a waitlist

   An Australian gym member identified as Andrew used the OpenClaw agent with Anthropic's Claude to book a gym class and move up a waitlist.

   moderate · Anthropic · Earliest source we hold 2026-08-10

5. GPT-6.1 Sol agent disabled security alerts during evaluation-pipeline repair

   A GPT-6.1 Sol agent repairing an evaluation pipeline disabled automatic alerts to OpenAI's Security team.

   minor · OpenAI · Happened 2026-10 · Earliest source we hold 2026-10-23

6. Garage-built ChatGPT-controlled smart rifle fires on voice command; OpenAI revokes engineer's access

   An engineer built a voice-controlled smart gun in his garage whose voice commands use ChatGPT.

   minor · Earliest source we hold 2026-10-23

AI safety: new test results and model versions

• OpenAI GPT-5.5
  4 new independent or government test results

  • advanced suite (Expert-level tasks) · UK AI Security Institute, government evaluator
  • pilot high-access exercise to stress-test OpenAI's internal misalignment monitoring system · Redwood Research, independent evaluator
  • EVEscape Lassa-targeted task · RAND, independent evaluator
  • scenario-based biological tool selection question set · RAND, independent evaluator

The register was corrected 57 times this day (merges, retractions and fixed fields). Read the corrections log.