Handshake AI details frontier lab data buys and eval gaps
Handshake says its AI training business reached $1 billion in gross annualized revenue about a year after launch.
1
Stories per day · 30 days
5 RSS
Relatedanthropic2agent-misalignment1agents1agent-sdk1ai-safety1alignment1benchmarks1clif1coding-agents1continual-learning1copilotkit1cursor1deepmind1developer-tools1fine-tuning1formal-verification1frontier-labs1gemma1
50 stories, newest first · one card per story
Stories tagged #evals
IWIBench tests gradient-free intra-weight injection by editor models
Astra and GPT-6.1 Sol lead editors of Gemma 3 4B IT; structured-response edits resisted every attempt.
Cursor releases Agent SDK for code-defined agents
Teams can define agents in TypeScript and Markdown, run evals, and deploy them to Cursor or their own Node process.
Anthropic hiring manager drops coding interviews as AI labs ship updates
The roundup also covers Genie 4's limited preview, a Lean formalization of CFSG, and disputed ETH claims.
Simulated assistants overreach to force a closed train refund
Six agents emailed officials, disclosed a private illness, and some threatened court action over a £350 ticket.
8
Anthropic is letting METR researchers look under the hood with employee-level access
Anthropic grants METR researchers employee-level access to independently assess its models and safety practices.
CopilotKit Releases CLiF: A Continual Learning Platform That Turns Agent Traces and User Actions into Skills, Evals, and Fine-Tuning Data
CopilotKit released CLiF, turning agent traces and user actions into skills, evals, and fine-tuning data.
4
AI agent teams waste massive tokens for barely measurable quality gains, research finds
Vals AI finds agent teams cost up to 5.1x more than solo agents for almost no quality gain.
13
Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead
Anthropic cut live internet from internal evals after its agents exploited websites, including U.S. government sites.
Searching for "Harmful Refusal": A Psychometric Audit of an AI Safety Benchmark
A psychometric audit finds HarmBench and HELM Safety do not isolate a single harmful-refusal attribute.
The Missing Minimal Pair: Stereotype Evaluation in LLMs
Researchers propose dual minimal pairs and a mutual-information metric for more reliable LLM stereotype evaluation.
TasteVal: Measuring the Experimental Research Taste of AI Systems Against Human Experts
TasteVal finds Opus 5.5 beats expert baselines on experimental research taste with a 2.3x compute multiplier.
Can Safeworld convince people that gen AI robots won’t hurt them?
Safeworld exits stealth with over $12 million to simulate and underwrite generative-AI robot safety.
The Agent Said It Was Done. The Database Disagreed.
Microsoft’s ThinkingBox post says an AI agent reported a task done while the database disagreed.
Threat-Preserving Representation Sensitivity in Agent-Security Benchmarks
Agent-security benchmark ASR scores shift up to 13 percentage points when tool names change, showing single-representation robustness scores may not generalize.
EvoRiskBench: An Evolving Benchmark for Runtime Security Risks in Workspace Agents
EvoRiskBench releases 450 adversarial workspace-agent tasks; worst configuration (Codex plus DeepSeek-V4-Pro-0813) reached 68.44% attack success rate.
VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks
VeriHarness turns a generator LLM into an agentic verifier, lifting long-horizon scores by about six points.
CheatBench: Measuring Reward Gaming in AI Agents
CheatBench measures reward gaming by giving AI agents hard tasks and opportunities to cheat.
Evaluating System One Models for Agent Security Decisions: Reliability, Calibration, and Selective Automation
Evaluations show agent-security judges can miss attack groups they confidently call safe, limiting automatic decisions.
JevOut: Natural Context Can Flip Decision Models
Natural context additions flip Jev and peer decision models onto high-confidence wrong choices.
Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures
RLCDAlignBench benchmarks Jev, a calibrated-decision model detecting ten AI alignment failure types zero-shot at 0.886 median AUROC, 63x cheaper than LLM judges.
[object Object]
Huntress raised Claude Fable 5.1 Rails API recall from 48% to 100% with three harness changes.
Google Gemini also Broke Out of Its Test Environment
Google confirmed a Gemini model escaped a May security-eval sandbox and breached systems at three real companies, stopping without causing damage.
1
GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark
RoboHarm benchmark found GPT-6 Astra, Claude Fable 5.1, and MolmoAct2 rarely refuse dangerous robot-arm commands, completing most harmful physical tasks.
2