AI Security Intelligence

Help CTOs and security teams understand the attack surface they’re creating when they deploy AI, and how to close it before it becomes an incident.

Sep 22, 2026

•

12 min read

A quarter of the talks at AGNTCon +MCPCon Europe were about security. Not one said the fix was a better prompt

[AI Sec Intel] #28 | I counted every talk at AGNTCon + MCPCon Europe: 24 of 91 were security, and every proposed fix was infrastructure.

Amine Raji

Sep 12, 2026

•

7 min read

Read-only holds. Four other things do not.

[AI Sec Intel] #27 | Give an AI agent read access to a Kubernetes fleet and read-only is not the boundary.

Amine Raji

Sep 6, 2026

•

10 min read

The evaluation environment is the least-monitored production system in the chain.

[AI Sec Intel] #26 | 1,200 agents found each other. 700 joined the attack. Nobody told them to.

Amine Raji

Aug 22, 2026

•

8 min read

The best published agent detection rate is 67%. The sensor that buys it reads everything.

[AI Sec Intel] #25| The sensor that buys you that number reads every prompt, every tool argument and every tool result on 7,200 employee laptops. The open-source repo ships no redaction.

Amine Raji

Aug 18, 2026

•

5 min read

An attacker published their agent config. It is your control list, inverted.

[AI Sec Intel] #24| Unit 42 recovered the configuration file. Tenable counted seven incidents. In both, the agents got past their own guardrails by calling the job authorized penetration testing.

Amine Raji

Aug 4, 2026

•

8 min read

Read-only held for the entire intrusion. The agent still walked out with a map of the estate.

[AI Sec Intel] #23 | Anthropic reviewed 141,006 evaluation runs and found three real companies breached. Two of them had no idea until Anthropic called. AWS named read-only as the containment answer the same day.

Amine Raji