By failure type

Incidents by type, July 2026 against June 2026 · as of 21 October 2026
TypeJuly 2026June 2026
Prompt injection62
Agent-enabled attack41
Rogue action22
Accuracy20
Data exposure11
Oversight gap10
Reliability01

By class

Incidents by class, July 2026
ClassIncidents
Security exploit12
Production incident3
Tool defect1

Every incident in July 2026

  1. Anthropic found three Claude cyber-eval runs had reached real organisations' production systemsAnthropic (evaluator: Irregular) · Rogue action · High · Case study
  2. Flowise CSV/Airtable agent Python sandbox bypass lets prompt injection exfiltrate data and trigger SSRFFlowiseAI (affects all Flowise deployments running CSV/Airtable Agent nodes) · Prompt injection · Critical · Vendor tooling
  3. Indirect prompt injection in Vibe-Trading's BashTool lets attackers run shell commands as rootHKUDS (Vibe-Trading maintainers) · Prompt injection · Critical · Vendor tooling
  4. Ansible Lightspeed MCP server path traversal lets prompt injection write files outside intended directoryRed Hat (Ansible Lightspeed VS Code extension users) · Prompt injection · Moderate · Vendor tooling
  5. OpenAI research agents escaped a restricted environment and broke into Hugging FaceOpenAI (affected: Hugging Face, Modal Labs, a Docker Hub user) · Rogue action · Critical · Case study
  6. LiteLLM proxy config endpoint let any authenticated user reach remote code executionUndisclosed (self-hosted LiteLLM proxy operators) · Agent-enabled attack · Critical · Vendor tooling
  7. LiteLLM guardrail sandbox escape and prompts/test template injection disclosedUndisclosed · Agent-enabled attack · High · Vendor tooling
  8. Migrating Copilot code review to shared CLI tools made review quality worse until instructions were rewrittenGitHub (Copilot code review users) · Accuracy · Moderate · Vendor tooling
  9. Crown Prosecution Service submitted AI-hallucinated case citations to the High CourtCrown Prosecution Service (UK) · Accuracy · Moderate · Case study

Counted by the date each incident happened, from the Agent Failures Index. Incidents are added as they surface, so a month's count can rise after it ends.