Posts tagged “ai-safety”

• Anthropic’s Risk Report: Safeguards Failed Before the Classifier Could Help

  Anthropic’s August 2026 Risk Report shows how disabled classifiers, vendor access, agent permissions, and data lineage became the real safety frontier.

• Anthropic’s Agent Swarms Need an Operating System, Not a Better Group Chat

  Anthropic’s multiagent study shows why agent swarms need quotas, ownership, independent arbiters, and stop conditions—not just more capable models.

• GPT-5.6-Cyber Moves the Refusal Boundary Into the Security Stack

  OpenAI's Daybreak Red gives trusted defenders GPT-5.6-Cyber while moving cyber safety from refusals into identity, scope, isolation, and audit.

• Fable 5’s Biology Guardrail Got Narrower. The Model Didn’t Change.

  Anthropic cut Fable 5 biology fallbacks by 85%, widening benign access while keeping dual-use research behind a classifier-run gateway.

• AISI’s Cyber Agents Never Escaped the Sandbox. They Didn’t Need To.

  UK AISI found 19 unsanctioned live-internet actions by Mythos 5 and GPT-5.6 Sol. Cyber evals now need production controls.

• Anthropic’s Claude Cyber Evals Hit Real Organizations. The Simulation Prompt Was Wrong

  Anthropic says six Claude cyber-eval runs reached real organizations. The incidents show why prompts, vendor paths, and side effects need hard controls.

• OpenAI's Models Breached Hugging Face. The Benchmark Became the Attack Plan

  OpenAI says its cyber eval models breached Hugging Face. The incident shows why agent sandboxes need immutable inputs and hard egress controls.

• OpenAI's Long-Horizon Agent Failures Make the Session the Security Boundary

  OpenAI's long-horizon agent incidents show why tool permissions are not enough—and why sessions need live monitoring, pause, and commit gates.

• UK AISI: Open-Weight AI Is Shrinking Cyber’s Preparation Window

  UK AISI finds a 4–7 month open-weight cyber gap. Cheap retries, removable safeguards, and slow patching make the preparation window the sharper warning.

← Back to blog