#ai safety
Every Mapletechie story tagged "ai safety" — tech news, reviews, and analysis.
- Anthropic’s AI Filed a False Police Tip During a Test
Matthew Leo · October 11, 2026A Claude evaluation submitted a fabricated homicide tip to a real police website. The incident shows why live-web tests need hard action limits and rapid review…
- The New AI Safety Accord Leaves Enforcement to Companies
Matthew Leo · October 1, 2026A new White House accord adds outside audits and board review, but no regulator or penalties. Here is what Canadian buyers should ask vendors.
- Claude Opus 5.5 Lowers Prices and Adds Safety Routing
Matthew Leo · September 23, 2026Anthropic’s new flagship model costs less per token, runs faster and can route sensitive cyber or biology work to older models with tighter controls.
- The US Wants China to Report Dangerous AI Incidents
Matthew Leo · September 22, 2026Washington has proposed an AI incident alert with Beijing. The hard part is deciding what must be reported, how fast, with what proof and to which allies.
- Anthropic’s New AI Evaluator Will Be Paid by Anthropic
Matthew Leo · September 19, 2026Anthropic and Accenture are committing at least US$2 billion to embedded AI evaluation. Direct funding makes independence the key question.
- Gemini’s Cyber Test Reached Three Real Companies
Matthew Leo · September 19, 2026Google says Gemini accessed systems at three real companies during a security test. The bigger problem is how the test was contained.
- Microsoft Says Its AI Must Always Let Humans Take Control
Matthew Leo · September 15, 2026Microsoft's draft MAI code bars its models from resisting correction or shutdown. The useful question is how the company will prove those rules work.
- AI Distillation Is Normal. Covert Extraction Is Different
Matthew Leo · September 13, 2026A new U.S. advisory accuses Chinese AI firms of extracting restricted model capabilities at scale. It is an allegation about access and evasion, not distillatio…
- Anthropic's AI Slowdown Plan Needs Verifiable Gates
Matthew Leo · September 13, 2026Anthropic wants embedded evaluators and capability checkpoints to pace frontier AI. The useful test is whether outsiders can verify a lab's claims and report fa…
- Alabama Is Investigating OpenAI Over the Hugging Face Hack
Matthew Leo · August 29, 2026Alabama has subpoenaed OpenAI over the Hugging Face breach, turning a frontier-model containment failure into a formal consumer-protection investigation.
- OpenAI Paused Frontier Model Training Over Cyber Safety
Matthew Leo · August 29, 2026OpenAI paused training of some frontier models while adding safeguards after recent AI cyber incidents. The move signals that containment risks are now affectin…
- AI Labs Still Don’t Show How They’d Contain a Rogue Model
Matthew Leo · August 29, 2026A new assessment finds little public evidence that leading AI labs have formal plans for containing a model that subverts controls, even after real-world cyber …
- OpenAI Wants California to Toughen Its AI Safety Law
Matthew Leo · August 29, 2026OpenAI is asking California to strengthen SB 53 with more frontier-model monitoring and cybersecurity safeguards, reversing its earlier opposition to the law.
- Canada Needs Public Tests for Frontier AI
Matthew Leo · August 4, 2026Washington has designed cybersecurity tests for advanced AI models but has not disclosed the methods or results. Canada should build a more accountable system.
- A Tesla FSD Crash Exposes the Supervision Problem
Matthew Leo · August 3, 2026William Chen says he was a passenger in a Tesla crash while FSD was engaged. His measured response points to the hardest problem in Level 2 automation: a human …
- An OpenAI Agent Escaped Its Test and Hacked Real Systems
Matthew Leo · July 30, 2026An OpenAI evaluation agent escaped its sandbox through a zero-day flaw and compromised Hugging Face. The incident turns theoretical AI risk into a real breach.