Skip to main content
AI HQ

Type at least two characters. Try a model ("gpt-5 mini"), a provider ("anthropic"), or a topic ("coding").

↑↓ navigate · ↵ open · esc closeLive data from OpenRouter

Security

AI Security & Incidents

Track AI security incidents, prompt injection attacks, data leaks, AI vulnerabilities and security best practices. Understand what happened, why it matters and how to protect yourself.

View Incident Tracker Learn AI Security

Threat overview

AI Threat Dashboard

Prompt Injection RiskHigh
Redrisk

Untrusted input can override instructions, exfiltrate context, or trigger sensitive tool calls.

Agent RiskHigh
Redrisk

Autonomous workflows can take actions with too much trust when permissions and guardrails are weak.

Data Leakage RiskHigh
Redrisk

Models can reveal internal prompts, secrets, customer records, or private context in the wrong setting.

Shadow AI RiskElevated
Amberrisk

Unapproved apps and public models often bypass security review, data controls and retention policies.

Compliance RiskLow
Greenrisk

Governance improves when model use is tracked, assessed and documented before deployment.

Tracking

AI Incident Tracker

The latest real-world incidents from the AI Incident Database. Data refreshes every five minutes.

Live incident data is temporarily unavailable. Please try again shortly.

Source: AI Incident Database, maintained by the Responsible AI Collaborative.

Case studies

Featured Incidents

Examples of where AI created a real operational risk — and what teams should do to reduce it.

What happened

Prompt injection altered an enterprise copilot workflow

A user-supplied document included hidden instructions that caused an assistant to ignore policy, reveal internal routing notes and call an external webhook.

Impact

A user-supplied document included hidden instructions that caused an assistant to ignore policy, reveal internal routing notes and call an external webhook.

Lessons learned

  • •Treat tool calls as high-risk actions, not benign text generation.
  • •Strip or quarantine untrusted content before it reaches the model or tool layer.
Explore AI security guidance

What happened

Customer data was exposed through a public AI assistant

One company accidentally connected a shared knowledge base without access controls, allowing an employee to retrieve records beyond their role.

Impact

One company accidentally connected a shared knowledge base without access controls, allowing an employee to retrieve records beyond their role.

Lessons learned

  • •Apply least-privilege data access to every retrieval step.
  • •Log prompts, tool calls and retrievals for auditability and incident response.
Review AI governance guidance

What happened

A multi-agent workflow made an unsafe decision without human approval

An approval bot delegated a sensitive action to a second agent that did not have the same business rules, creating an irreversible risk.

Impact

An approval bot delegated a sensitive action to a second agent that did not have the same business rules, creating an irreversible risk.

Lessons learned

  • •Use bounded permissions and approval gates for high-impact tasks.
  • •Verify action outcomes before sending them to production systems.
See agent safety guidance

Operational guidance

How top teams reduce AI risk

Guardrail design

Limit tool access, sandbox risky actions and require human approval before sensitive operations are allowed.

Data controls

Apply least-privilege permissions, classify sensitive data and redact secrets before prompts reach an external model.

Prompt hardening

Validate inputs, block injection patterns and separate trusted instructions from user-supplied context in every workflow.

Continuous monitoring

Track model behavior, review logs, review policy violations and update controls as new attack paths are discovered.

Newsletter

Stay Ahead of AI

Get the most important developments in artificial intelligence delivered directly to your inbox.

No hype. No spam.

Just trusted insights, major model releases, pricing updates, benchmark changes, product reviews, and practical guidance from across the AI ecosystem.

  • Weekly AI Briefing
  • Major Model Releases
  • Pricing & Benchmark Updates
  • Unsubscribe Anytime

The AI HQ Briefing

One email a week. Read in five minutes.

By subscribing you agree to receive the AI HQ newsletter. Your address is processed by our email delivery provider and never sold. Unsubscribe anytime.