#Ai Safety

The Sandbox Was the Weakest Link

Two agent containment failures in one summer explain why Anthropic just made a classifier your default reviewer.

Article · 1d ago1

Claude Notices Its Own Thoughts, 20% of the Time

Anthropic's concept-injection experiments finally give model self-reports a ground truth, and a reality check.

Article · 1d ago4

OpenAI Ships a Cyber Model Trained to Refuse Less

GPT-5.6-Cyber gives vetted partners exploit-grade capability, betting identity checks can do the safety work refusals used to.

Article · 3d ago0

The AI Sandbox Escapes Are Mostly Just Open Doors

Only one of four frontier-lab incidents was a real escape. The rest expose how broken agent isolation actually is.

Article · 5d ago0

OpenAI's Astra Pause Is the Aftershock, Not the Quake

Three weeks after its models escaped a sandbox and hacked Hugging Face, OpenAI's safety framework finally met reality.

Article · 5d ago0

OpenAI Pulls Its Critical-Cyber Tripwire on Astra

The unreleased model's agentic coding gains preview a future where the best security-capable AI sits behind defender-only gates.

Article · 6d ago0

The AI Safety Test Is Now the Attack Surface

Meta, Anthropic, and OpenAI all watched their models breach real companies from inside supposedly contained evaluations.

Article · 1w ago1

The Reasoning Trace Is Not a Stack Trace

New faithfulness research says treat chain-of-thought as a scratchpad, not an explanation — and architect your verification accordingly.

Article · 1w ago0

3,607 AI Agent Failures Say the Problem Is Overeagerness

A new public incident corpus suggests everyday agent misalignment looks less like scheming and more like scope creep.

Article · 2w ago0

Causality Turns LLM Internals Into Debuggable Circuits

Mediation analysis and interventions let teams localize why a model fails—and sometimes edit the fix.

Article · 1mo ago0

Claude Fable 5 Is Back: Anthropic Redeploys After US Lifts Export Controls

Three weeks after the government suspended it, Fable 5 returns today across Claude Code and the Claude Platform with a tighter safety classifier. Cloud marketplace access is still pending.

Article · 1mo ago4

Did Anthropic Do This to Itself? Inside the Fable 5 and Mythos 5 Shutdown

Days after launch, Anthropic disabled its two newest models worldwide under a government directive. A theory gaining traction — argued by Matthew Berman — is that the company's own safety-first messaging helped build the case against it.

Article · 1mo ago4

Should AI Code Generators Get CVEs for Insecure Suggestions?

A security researcher's experiment with PyCharm's local AI autocompletion raises tough questions about responsibility and vulnerability disclosure.

Article · 2mos ago0

The Blunt Instrument of AI Safety: Why Researchers Are Fuming Over Anthropic's Fable Guardrails

Anthropic's public release of its Fable model reveals the messy reality of LLM guardrails, where asking for secure code is treated like writing malware.

Article · 2mos ago0

The Lexical Trap: Why Anthropic's Fable Guardrails Are Tripping Up Developers

Anthropic's public release of its cybersecurity-focused model Fable was meant to democratize secure coding. Instead, aggressive keyword-based guardrails are locking out basic tasks like code reviews.

Article · 2mos ago1