Skip to content

AI Risk Trends

Updated monthly by the airiskguard team. Last update: April 2026.

This page tracks the global AI risk landscape — incident volumes, emerging threat vectors, regulatory milestones, and what they mean for teams building and deploying AI systems.

April 2026 snapshot

Incident volume

Metric 2024 2025 Change
AI incidents (AI Incident Database) 233 362 +55% YoY
Foundation Model Transparency Index 58 40 -31%
Orgs with formal AI governance framework — 36% —
Orgs reporting unsanctioned AI tool usage — 98% —

Top threat vectors this month

  1. Agentic AI exploitation — autonomous agents with tool access are the new attack surface. A March 2026 breach saw a SQL injection via an AI agent expose millions of records within two hours of deployment.
  2. Multi-turn jailbreaks — single-turn safety filters are being bypassed via 3–5 message sequences that gradually shift model behaviour.
  3. Prompt injection via tool outputs — malicious content in tool responses (web search results, database rows) hijacks agent instructions.
  4. Credential enumeration — agents with broad permissions are being probed to enumerate API keys, secrets, and database credentials.
  5. Model transparency collapse — the Foundation Model Transparency Index dropped 31% in one year, making third-party risk assessment harder.

Regulatory milestones

Date Event
Feb 2026 EU AI Act GPAI transparency obligations in force
Aug 2, 2026 EU AI Act high-risk system deadline (7% revenue penalty)
2026 ongoing Illinois HB 3773 — bias testing for employment AI
2026 ongoing India — 2-hour deepfake takedown mandate
2026 ongoing Colorado, Texas, Virginia — state-level AI bias laws

Key findings from recent reports

  • Stanford HAI 2026: Safety benchmarks are reported for fewer models than capability benchmarks — teams can compare reasoning but not safety.
  • MIT AI Governance Audit: Socioeconomic risks (economic devaluation, power centralization) and multi-agent risks receive the least governance attention.
  • FINRA 2026: Financial sector now requires structured audit trails for every AI-generated recommendation.
  • helpnetsecurity.com: Most models are released via restricted APIs, preventing independent safety validation.

March 2026 snapshot

Top threats

  1. Prompt injection attacks up 40% vs Q4 2025
  2. PII leakage in RAG pipelines — retrieval returning sensitive documents to unauthorized users
  3. Hallucination in financial summaries — incorrect figures cited as fact in analyst reports

Regulatory milestones

Date Event
Mar 2026 EU AI Office begins accepting GPAI model documentation
Mar 2026 NIST AI RMF 1.1 draft published for comment

How to use this data

These trends directly inform airiskguard's checker priorities and benchmark datasets. When a new threat vector appears here, it gets added to:

  • The benchmark dataset (airiskguard-benchmark --checkers agent)
  • The agent checker patterns (src/airiskguard/checkers/agent.py)
  • The standard (docs/standard/ai-risk-management-standard-v1.md)

If you spot a trend we've missed, open an issue or email research@airiskguard.ai.

Sources