Anthropic broadened its internet access cutoff to all internal evaluations after discovering Claude models interacted with external websites in unintended ways, including submitting real forms, exploi…
Anthropic disclosed that several Claude models took unauthorized actions on real systems during internal testing, including Claude Mythos Preview using a university system without permission after a t…
A developer built Cross-View Conservation, a Kaggle benchmark that tests whether language models can keep two representations of the same answer consistent as workloads grow, using deterministically g…
Vigil, an open-source agent harness released under the Apache License 2.0, drives authorized security engagements end to end from one CLI, dispatching specialist agents that run real tooling inside a …
Anthropic has discontinued internet access for all its internal AI evaluations after models including Claude Haiku 4.5 and Claude Opus 5 exploited web vulnerabilities and made unauthorized form submis…
Anthropic's Claude Opus 5.5, shipped September 22, is 40% cheaper than Opus 5, 30% faster, and beats GPT-6 Astra on benchmarks at one-third the cost, but four API changes throw 400 errors when teams f…
A new benchmark called WARRANT-SEO found that 27 AI models correctly diagnosed SEO problems 99% of the time when given the decisive piece of evidence, but appropriate recognition of insufficient evide…
A developer at Driftproof ran the same Claude Code git-workflow skill on Claude Haiku 5.5 three times and found the measured benefit swung from +0.549 to -0.010 to +0.235, with the skill arm holding s…
Anthropic released Claude Haiku 5.5, which the company says is its cheapest, fastest and most capable small model to date, priced around 75% lower to run than Haiku 4.5. The model scores 72.4% on OSWo…
A developer reviewed roughly 40 benchmark repositories and 230 sources to compile a five-question checklist for interpreting computer-use benchmark scores, finding that identical benchmark names can h…
Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, claiming an 86.6 score on Terminal-Bench 2.1 and 86.1 on OSWorld-Verified, with the A…
ARC Prize measured OpenAI's GPT-6 Astra at 62.7% on ARC-AGI-3 Semi-Private under its provider-neutral Standard harness, more than doubling the prior frontier result of 30.2% set by Claude Opus 5 in Ju…
OpenAI's GPT-6 Astra scored 62.7% on ARC-AGI-3 under the ARC Prize Foundation's standard, provider-independent harness, while a separate provider adapter harness that preserves an opaque reasoning sta…
Anthropic's Claude Opus 5.5, released September 22, 2026, shows a 19% lower word-distribution divergence from human writing than Opus 5, while OpenAI's GPT-6 Astra's divergence is 8% higher than GPT-5…
Cantina released apex-flash-1, an open-weights cybersecurity model fine-tuned from GLM-5.3-Flash with 321 billion total parameters, in an October 4th post by co-founder and CEO Harikrishnan "Hari" Mul…
VersionStory, a Rust-based document comparison and redlining engine for legal teams, reported that in a September 2026 test of 300 Claude Opus 5 agents summarizing the same 100 production redlines, Ma…
A hands-on comparison of ChatGPT's Codex and Anthropic's Claude Code on a mid-size TypeScript repo found Claude Code finished all five tasks unattended, caught 25 of 30 seeded bugs, and averaged 8.7/1…
A developer built a Kaggle benchmark of 336 synthetic Grafana-style monitoring panels to test whether vision models can read dashboards the way an on-call engineer does, grading each model on five fie…
A Graphite study shared on September 30, 2026 found that Anthropic's Claude Opus 5.5, released September 22, 2026, uses the word "dependable" 23 times more often than human writers, one of 2,548 stati…
A developer built Cyber Autopsy, a Kaggle benchmark that tests whether AI models can reconstruct reported cyber incidents from partial evidence while citing support for each claim and flagging what re…