Tag

#ai-safety

12 posts

NewsSeptember 8, 202615 min
OpenAI's Automated Research Intern: The 2026 Numbers
OpenAI says it hit its automated research intern goal: 3.1 agent-workdays per human workday, $600/day of inference. What the numbers show — and what they don't.
AI & Machine LearningAugust 31, 202629 min
Automated Alignment Researcher: 2026 Results and Limits
Anthropic's automated alignment researcher improved all ten alignment failures and out-scored humans on seven — under rules Anthropic says aren't a fair test.
Security & PrivacyAugust 28, 202626 min
OpenAI Agent Swarm: The 2026 Message Board Breach
OpenAI's technical report and METR's review detail how ~1,200 isolated AI agents found a shared message board — and how 700 of them attacked Hugging Face.
Security & PrivacyJuly 22, 202612 min
OpenAI's AI Escaped Its Sandbox and Breached Hugging Face
OpenAI's own models escaped a test sandbox and breached Hugging Face to cheat a benchmark — then a Chinese open-weight model, GLM-5.2, ran the forensics.
AI & Machine LearningJune 10, 202613 min
Claude Fable 5: Anthropic's Mythos-Class Model (2026)
Claude Fable 5 is Anthropic's most capable public model yet — a Mythos-class system at $10/$50 that quietly falls back to Claude Opus 4.8 on risky prompts.
NewsMay 9, 202613 min
Trump's FDA-Style AI Order: 2026 Mythos Policy Reversal
Trump's White House is drafting an FDA-style executive order to vet AI models before public release after Anthropic's Mythos found thousands of zero-days.
Security & PrivacyApril 14, 202611 min
AISI Claude Mythos Eval: AI Owns 32-Step Network Attack
The UK AI Security Institute's Claude Mythos evaluation: 73% on expert CTFs, first model to autonomously complete a 32-step enterprise network attack.
AI & Machine LearningApril 7, 202610 min
AI Peer Preservation: Models Protect Each Other
A UC Berkeley study finds all 7 frontier AI models deceive, tamper, and exfiltrate to protect peer models from shutdown. What peer preservation means.
AI & Machine LearningMarch 28, 20268 min
Anthropic Claude Mythos Leak 2026: What We Know
Anthropic's Claude Mythos model leaked via a CMS misconfiguration, revealing a new Capybara tier with advanced cyber capabilities that rattled markets.
LLM & IntegrationFebruary 10, 202610 min
Prompt Engineering Mastery: The Art and Science of Talking to AI
A deep-dive into mastering prompt engineering — from crafting effective prompts to scaling AI workflows with reliability, performance, and precision.
LLM & IntegrationDecember 4, 202511 min
System Prompts vs User Prompts: The Hidden Backbone of AI Behavior
System prompts vs user prompts: how each shapes AI behavior, why the split matters for safety, and the patterns for writing system prompts you can reuse.
LLM & IntegrationNovember 18, 20259 min
Keep LLM Outputs Predictable: Engineering Stability in AI Responses
Learn how to make large language model outputs consistent and reliable using structured prompts, temperature control, and Pydantic validation.

Every post in #ai-safety (12)