AI Research & Models
Joe Benton, who led a safety research team at Anthropic, and Josh Engels, formerly of Google DeepMind, left their posts and are joining AI-incident nonprofit METR, days after ex-Anthropic researcher Jacob Coxon's viral resignation post reignited Washington's AI-safety debate.
why it mattersThe exodus is already fueling two new bills in Congress, the FRONTIER Act and the Ban Artificial Superintelligence Act.
Anthropic's latest Claude Code release adds a `claude plugin eval` command that scores a plugin's test suite and returns reproducible JSON/HTML reports, plus `/output-style` for switching styles mid-session across CLI, VS Code, web, and Slack.
why it mattersPlugin authors get a real test harness instead of manual spot-checks.
Codex's model picker and Amazon Bedrock catalog now include GPT-6 Astra, while the Codex Python SDK (0.154.0) adds 'max' and 'ultra' reasoning-effort values and better session-resume history controls.
why it mattersSignals OpenAI is pushing its frontier model deeper into enterprise cloud deployments of Codex.
DeepSeek's new 552B-parameter mixture-of-experts model adds native multimodal vision and a redesigned encoder-decoder architecture; the company says independent tests put it ahead of V4-Pro on performance, cost and speed, and will reroute all V4-Pro API traffic to it starting September 14.
why it mattersA smaller model beating its own flagship undercuts the assumption that bigger is always better.
Sakana's multi-agent orchestration model routes queries across a pool of underlying models behind one API; version 2.0 drops Claude Fable 5/5.1 and GPT-6 Astra from that pool while still leading published benchmarks like Chartography and DeepSWE.
why it mattersSakana is betting customers care more about not depending on one proprietary frontier model than about using the single best one.
The latest Docker Desktop release fixes CVE-2026-17106, a destination-escape bug in `docker container cp`, and adds the ability to start and stop Kubernetes cloud clusters directly from Docker Offload.
why it mattersThe CVE affects anyone copying files out of untrusted containers.
Both releases pull in the OpenSSL 3.5.8 security upgrade; the LTS line adds a STORE loader for hardware security modules, refreshed NSS root certificates, and speedups to URL parsing and HTTP responses.
After a bug in Liquid's Elements software let an attacker mint roughly 4,000 BTC of unbacked LBTC and peg it out on September 6, the attacker returned about 3,400 BTC but is holding the remaining ~598.5 BTC (~$47M) for a 10% bounty. Blockstream says it won't pay, calling the demand theft, and plans to work with law enforcement and forensic firms to trace the funds.
why it mattersA hard line against paying hackers could set precedent for how other protocols handle post-hack extortion.
US spot Bitcoin ETFs shed $282.6M on Thursday, led by ARK 21Shares' ARKB ($164M), Grayscale's GBTC ($36M) and Fidelity's FBTC ($33.6M), bringing the three-day outflow total to $449M even as cumulative net inflows remain above $55B.
Price is 1.20× the 200-week average. Weekly closes, USD, as of 28 October 2026 · data blockchain.com.