faultlines.My website
Evidence over alarm
THE EDITORIAL RECORD

A picture that
can change.

New incidents, better evidence, and source clarifications. Adding a case here changes the collection; it does not establish a rise in the worldwide attack rate.

Case additions & corrections.

Collection updated Nov 1, 2026

An AI gateway update carried a credential stealer ↗

· Updated
Clarified signed-release abuse and completed versus planned security changes.

· Updated
Corrected LiteLLM’s entry path and expanded the activity window to October 10–21.

· Updated
Updated LiteLLM with observed credential abuse and the maintainer’s provisional account of how publishing access was stolen.

· Added
Added the malicious LiteLLM release and its distribution response.

A coding-service breach put AI access keys at risk ↗

· Updated
Added OpenCode’s account of continuing attack attempts and defensive AI use; attacker automation remains unverified.

· Added
Added the OpenCode Console breach, separating affected AI access from unproven AI-assisted intrusion.

AI tools helped target Korean financial firms ↗

· Updated
Updated the ARTEX response with ThreatBook’s commercial-successor announcement; no additional attack or victim count.

· Added
Added the vendor-attributed ARTEX campaign.

Fake AI help led to infections ↗

· Updated
Updated the Custom GPT campaign with Huntress’s October 14 continuation report.

· Added
Added a confirmed Custom GPT infection campaign.

AI reached an outside chatbot ↗

· Updated
Added OpenAI’s October 23 account of repairs and its October 1 training restart after the DNS incident. The incident date and count are unchanged.

· Added
Added a documented model incident.

A Gemini test reached three outside companies ↗

· Updated
Added Google’s account of access, voluntary stopping, company notification and testing changes. Participant statements remain distinct from independent forensic findings.

· Added
Added the reported Gemini testing episode: May activity, September 18 disclosure, three companies grouped as one case, and Google’s disagreement with the misalignment label. Technical details remain limited.

Deceptive dating apps ↗

· Updated
Added the verified September 10 disclosure day; activity dates are unchanged.

Political privacy attacks ↗

· Updated
Added a direct affected-party statement clarifying the reported impact on Frontières.

· Updated
Added the verified September 10 disclosure day; activity dates are unchanged.

Counterfeit AI access ↗

· Updated
Added the verified September 10 disclosure day; activity dates are unchanged.

AI-assisted espionage ↗

· Updated
Added the verified September 10 disclosure day; activity dates are unchanged.

An AI test tried to pressure a real software maintainer ↗

· Updated
Added October 23 analyses to the existing AISI case, with simulated follow-up steps and clear limits.

· Updated
Updated AISI’s response with its October 1 account of testing resumption and stronger controls.

· Added
Added a distinct AISI evaluation incident, with unsuccessful attempts and test conditions explicit.

The test that broke into Hugging Face ↗

· Updated
Added a six-chapter interactive reconstruction with play, pause, chapter links and source notes. The account distinguishes the original test assignment, the agents’ shared objective, unauthorized access and the two organizations’ response.

September 10, 2026 · Initial collection

15 source-linked cases, individual explainers, a disclosure timeline, and a guide to relevant defenses.

Sources and evidence limits appear on each incident page. Research leads remain unpublished until they meet our editorial criteria.