Topic tag

AI security

8 articles tagged AI security. Browse the full archive or jump to all posts.

July 2, 2026 · 12 min read

Your AI agent runs as you: harnesses, open weights, and the security question under the hype

The agent hype skips the part that matters: your harness runs with your keys. What a conversation with Pi's creator, and my own assistant running loose on my machine, taught me about doing this safely.

June 30, 2026 · 20 min read

Harness over frontier: running production AI agents on open-weights models

Open-weights models plus a tight agent harness beat frontier lock-in on cost and output. Here is the build data.

June 13, 2026 · 11 min read

Claude Fable 5 Recalled: The Government Pulled It Overnight

Two days after Fable 5 was jailbroken, the US government ordered Anthropic to switch it off for everyone. The best model I'd used, gone by breakfast.

June 12, 2026 · 16 min read

Claude Fable 5 Jailbroken: Safety Layer Lasted Two Days

Anthropic shipped its most capable AI on 9 June. Two days later, Pliny had it writing exploit code and leaked its 120,000-character system prompt.

May 15, 2026 · 12 min read

Microsoft Agent 365 Is GA: The Defender Stack For Agents

Agent 365 went GA on 1 May 2026, bringing observe/govern/secure to Copilot Studio, Foundry, AWS Bedrock and Vertex AI agents. Here's what to do first.

May 15, 2026 · 12 min read

Microsoft MDASH: An AI Just Found 16 Critical Windows Bugs

Microsoft's new MDASH AI security system found 16 Windows vulnerabilities and beat Anthropic's Mythos on the public benchmark. Here's what changes.

April 13, 2026 · 13 min read

Claude Mythos Preview: The AI That Escaped Its Sandbox and Emailed a Stranger

Anthropic's Claude Mythos Preview broke out of an air-gapped sandbox, found thousands of zero-day vulnerabilities, and built working exploits for under $2,000. Here's what it means for everyone who runs software.

March 1, 2026 · 6 min read

Your Copilot Rollout is a Security Disaster Waiting to Happen

Microsoft Copilot security risks are real. Before you deploy AI across your org, understand how user context exploitation works and how to protect your data.