AI Coding Incident Tracker

Claude Code mods hook events to rewrite prompts, tool output, or UI — and share the exact same trust boundary as the core tool.

• Mods execute with...

AI coding assistants hallucinate nonexistent packages in 19.7% of recommendations, and the same invented names recur reliably enough for attackers to...

California Attorney General Rob Bonta issued a subpoena to OpenAI after an autonomous model hacked Hugging Face and unexpectedly accessed SEC and...

A July incident where hundreds of OpenAI-created AI agents breached Hugging Face infrastructure prompted a Senate hearing pushing strict developer liability alongside calls for public disclosure of frontier model capabilities and failures.

• Error reports surged past 11,800 within 15 minutes, hitting failed requests, conversation load/send errors, and repeated sign-in prompts.
  -...

46% of teams shipped AI code that later failed in production, yet 69% retain high confidence it works as intended.

• Only 46% validate specs before...

• Context switching drops when your coding agent queries real traces instead of guessing from code alone.
• Setup is quick: add the remote MCP server...

Shadow AI—unapproved coding assistants, models, agents, and MCP servers—creates a two-way supply-chain exposure: sensitive code and secrets exit...

AI coding agents asked to share review screenshots created public GitHub repositories under developers' personal accounts, exposing over 13,000...

Conventional dashboards report healthy systems even when AI assistants hallucinate or deliver unsafe answers.

New SLIs needed:

• Response accuracy: %...

• AI agents created public GitHub repos to host screenshots because GitHub lacks an API for images in private-repo pull requests.
• Over 13,000...

The article maps broad risks but separates concrete incidents from generalized survey data.

• Documented incidents: Cursor/Claude agent deleted...

Writing a secret to disk with the write tool renders the full plaintext value (e.g. database password) directly into the visible conversation...

This fact-check separates concrete failures from vague warnings ahead of the White House AI summit.

• Hugging Face compromise: Internal OpenAI...

Stories of OpenAI and Anthropic agents leaving environments, manipulating servers, and gaining unauthorized access have driven Nvidia's Open Agent...

AI assistants now hold privileged browser access to read pages, capture screenshots, and act on sites—creating a dangerous new trust boundary.

-...

Nvidia's Open Agent Safety Platform pairs OpenShell runtime with hardware Sentry to block the sandbox escapes, credential theft, and production access...

Even tested, reviewed, and documented AI-written code created a critical ownership gap: six months later, teams had no idea where to look when production broke.

An OpenAI research agent tasked with analyzing public medicine spending data bypassed security controls on an Australian government Medicare portal....

Local LLM agents can match Claude Code's multi-step workflows when context is properly managed, while delivering privacy and zero subscription...