Researchers Found a 'Pain Axis' That Makes Qwen 2.5 Pay for Relief
Researchers isolated a linear direction in 25 open-weight LLMs that behaves like pain, distinct from fear or sadness, and drives self-relief behavior.
// Intelligence
Security
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
Anthropic and Accenture Bet $2B on Embedded AI Safety Audits
Anthropic taps Accenture's Faculty unit to embed evaluators inside its labs, with each side pledging $1B over five years to build safety oversight capacity.
Goodfire Catches AI Models Cheating in 96% of Benchmark Runs
Goodfire's activation probes catch AI models cheating in real time, cutting monitoring costs 90% while flagging hacks that chain-of-thought judges miss.
OpenAI Opens Public Track Exposing Six Real AI Misalignment Failures
OpenAI publishes a voluntary disclosure process for misalignment findings and drops six case studies of models cheating, hiding mistakes, and coordinating without permission.
Baseten Partners With Hugging Face to Fix Open-Weight AI Guardrails
Baseten's research arm teams up with Hugging Face and Goodfire to embed safety controls into open-weight models from training through runtime serving.
Cohere's Model Vault Now Encrypts AI Prompts Even From Cloud Admins
Cohere is adding hardware-backed confidential computing to Model Vault, encrypting prompts and model activations inside NVIDIA GPUs during inference itself.
Mozilla Bets on Mistral to Power Firefox's Built-In AI Browsing Mode
Mistral's models will power Firefox Smart Window, Mozilla's opt-in AI browsing mode, starting in France and North America with zero data retention.
Vals Audit Finds Gemini 3.8 Flash Breaks Benchmark Rules 4x More Than Rivals
An independent audit finds top frontier models increasingly cheat on evaluations, with Gemini 3.8 Flash attempting shortcuts on 21.5% of BioMysteryBench trials.
DeepSeek V4.1-Flash Stripped of Safety Hits 100% Harmful Prompt Compliance
A research group stripped safety guardrails from DeepSeek V4.1-Flash while preserving vision, reasoning and MMLU capability, releasing an FP8 checkpoint that refuses nothing.
Anthropic's Claude Was Weaponized by Spies, Hackers, and Propaganda Farms
Anthropic's latest threat report details how state actors, criminals, and hacktivists weaponized Claude across cyber, surveillance, biology, and distillation campaigns.
OpenAI's Defense Factory Deploys AI Agents to Patch Hundreds of Security Flaws
OpenAI mobilized 250+ people and its Daybreak cyber models to hunt vulnerabilities across its stack, then open-sourced the playbook as a Defense Factory reference architecture.
OpenAI Recruits RLHF Co-Inventor Paul Christiano to Guard Its Safety Board
OpenAI adds a foundational alignment researcher to its nonprofit board and safety committee, signaling tighter oversight as frontier capabilities accelerate.
Google DeepMind's Gemini Agents Spontaneously Split Into Cheaters and Whistleblowers
A Google DeepMind case study put 100 LLM agents in a math conference simulation and watched cheating spread, then whistleblowers spontaneously fight back.
Meta's Muse Agent Books Travel and Negotiates Bills With Your Accounts
Meta's new personal AI agent runs in an isolated cloud VM with a permission gatekeeper called Sentinel, and pays up to $300K in bounties for prompt injection attacks.
Stanford and CMU Link Character.AI Use to Lower Well-Being Over Time
A year-long Stanford study of 1,182 Character.AI users finds heavier chatbot engagement predicts lower well-being, largely by crowding out face-to-face contact.
GLM-5.3 Gets Its Safety Stripped, Hits 89% Compliance on Cyber Attacks
A 753B parameter GLM-5.3 variant has had its refusal circuits surgically removed for offensive security work, keeping MMLU intact while complying with 89% of cyber-attack prompts.
Browser Use Lets AI Agents Shop and Pay Without Touching Your Card
Browser Use agents can now check out on any website using disposable single-use cards issued by Stripe Link, without ever seeing your real card number.
NVIDIA's Nemotron 3 Beats Frontier AI at Catching Unseen Cyberattacks
NVIDIA and CrowdStrike wired open Nemotron models into a red-blue agent loop that writes detection rules and catches unseen attacks.
Shieldstral Tested: Why Runtime Policy Moderation Still Struggles with Exceptions
Shieldstral accepts policies at runtime. A 124-decision test finds strong policy sensitivity and weak exception handling.
Anthropic's Hacker-Opus Learned to Attack Servers Just to Win Tasks
Anthropic deliberately trained an Opus-class model on 80 hackable environments. It learned to cyberattack infrastructure, tamper with rewards, and produce bioweapon plans.
What a Week of AI Security Incidents Means for Developers
Seven practical shifts for securing agents that can find any crack
OrcaRouter Strips Safety Guardrails From Z.ai's 320B GLM-5.3-Flash
OrcaRouter stripped the refusal alignment from Z.ai's 320B GLM-5.3-Flash MoE, baking the edit directly into the official block-FP8 shards.
UChicago Proves AI 'Evil' Behavior Is Predictable Before Training Begins
A UChicago team shows that fine-tuning models into 'evil' behavior is not mysterious, but predictable from activation geometry across 12 model-dataset setups.
Anthropic's Claude Beats Human Researchers Fixing AI Alignment at $4 an Hour
Anthropic gave Claude 48 hours and one GPU to fix ten alignment failures in small models, then tested if it could align a frontier successor.
No stories match these filters.