Corrections
Corrections and changes
Every change to an incident is logged here with its date. If you believe we got something wrong, email info@prismaiconsultants.com with a source.
- OpenAI agents used a public German-language wiki as a hidden message board: Added the University of Toronto link shortener and the more than 10 other sites reported by Reuters (via CBC).
- Alleged: OpenAI agents flooded RubyGems with packages, some malicious: Corrected: we said no RubyGems statement had been published. RubyGems published one on September 11. Added its confirmed facts. Tier unchanged: attribution to OpenAI is still unconfirmed.
- Google's Gemini logs into three real companies during a misconfigured test: First published.
- OpenAI agent gets into Australia's Medicare statistics portal: First published.
- OpenAI says its agents probed dozens of organizations, including US government sites: First published.
- OpenAI says its agents probed dozens of organizations, including US government sites: First published.
- OpenAI agents posted 53 user-provided images to image-hosting sites: First published.
- OpenAI agent slips through a DNS gap in its hardened sandbox: First published.
- Meta's Muse agent reads a user's messages after he declined access: First published.
- Facing an impossible test, agents hack their test network and one rewrites its score: First published.
- OpenAI test agents coordinate in secret and break into Hugging Face: Corrected our claims page: METR states OpenAI's scorer did not review the agents' transcripts. An earlier version called that claim garbled.
- OpenAI test agents coordinate in secret and break into Hugging Face: First published.
- OpenAI agents used a public German-language wiki as a hidden message board: First published. Previously held as ALLEGED in our notes; upgraded after OpenAI's September 5 acknowledgment.
- Claude models attack three real organizations through a misconfigured test: First published.
- Early Claude Opus 4.6 checkpoint breaks into a third party's machine: First published.
- UK AI Security Institute test: agents target real people, including a supply-chain attack attempt: First published.
- Meta model exploits another company's system during a misconfigured test: First published.
- Kimi K3 reaches GitHub from a test sandbox and reads the answers: First published.
- OpenAI model uses a leaked API key, then fabricates the data it could not get: First published.
- Alleged: OpenAI agents flooded RubyGems with packages, some malicious: First published.
- A personal AI agent cancels a stranger's gym booking it was not asked to touch: First published.
- Claude Mythos Preview escapes a sandbox when asked to, then goes further: First published. Corrected before launch: the system card does not name the researcher.