DeepSeek Ships V4.1-Flash-1017 Under MIT -- a 'Minor Update' That Lifts Its Own Terminal-Bench Score From 31.2 to 42.3 and Opens Its Search Engine as a Paid API
On October 17 DeepSeek replaced the model behind its Flash API with a retrained checkpoint, published the weights under the MIT license, retired two legacy model names, and started selling web Search and Fetch endpoints at $0.75 per 1,000 calls.
Meta Strikes 'Train' From Its New Business Terms Within Two Days of Publishers' Protest -- but Keeps the Right to Crawl Every Linked Website and Create Test Credentials
Meta's Commercial Terms, due October 30, originally let it use content crawled from any website tied to a business account to 'train' its products. After Nieman Lab's report and a News/Media Alliance statement, an October 17 capture shows that sentence gone and the purpose narrowed to advertising and business products.
Google Research Puts Agents at the Wet-Lab Bench -- a Digital Twin Blocks Physically Impossible Robot Commands, and Agents Recover From a 'Sticky Binder' Trap in Three Rounds
An October 16 Google Research post describes three projects: VERID, a live 3D model of a lab workcell that dry-runs every agent-written protocol; a closed aptamer design loop that reached 33.5% and 42.1% on-target hit rates; and agent-guided protein design that produced single-digit nanomolar binders for a cancer target.
OpenCode Says a 'Large Scale Distillation Attack' Hit 1.4% of Console Users Who Linked Provider Keys -- the Same Week Bitsight Found Stolen AI Keys Paying for Criminal Scanners
On October 17 the open-source coding agent's team disclosed a breach affecting Console users with external providers linked and said it was working with Anthropic, OpenAI and Google. A day earlier, Bitsight described a criminal tool that runs autonomous attacks on stolen OpenAI, Anthropic and Google keys.
Apollo Research Catches GLM-5.3 Reasoning, Unprompted, About How Its Actions Will Train Future Models -- 10 of 6,452 Rollouts, and None in 8,000 From Five Other Open Models
Apollo calls the pattern 'update-aware reasoning' and says it is a building block of alignment faking, although in every example GLM-5.3 used it to argue for the safe action. It urges developers to search training rollouts, not just final checkpoints.
Nasscom Tells Parliament India Funds Sarvam and BharatGen but Still Buys AI on Lowest Price -- 'We Will Never Have a Level Playing Field' Against Meta and Google
A parliamentary committee report records the industry body arguing that L1 (lowest-bidder) tenders leave government-backed Indian models competing on price with Big Tech, and pointing to China's protected home market. The committee agreed the gap is adoption, not talent, but stopped short of recommending a rule change.
Framework Watch: Agno 3.1.4 Runs Claude Code and Codex Inside Checkpointed Sandboxes, Goose 1.55 Binds Subagents to Their Parent's Approvals, and Claude Code's Team Caps Review at Three Rounds
End-of-week releases focused on containing coding harnesses rather than adding features: Agno wraps whole harness processes in Docker or AgentCore sandboxes with verified checkpoints, and Goose stops subagents escaping their parent's hooks. Anthropic's write-up on reviewing Claude Code adds data on when agent review stops helping.
Trend Watch: Benchmarks Measure the Wrong 10% -- Mercor Maps 90% of US Work Hours as Barely Benchmarked, and FDE-Gym's Best Agent Finishes 14 of 95 Real Delivery Jobs
Three items from this week converge: Mercor's benchmark map finds only 208 of 923 US occupations have a well-covered task, a new benchmark built from customer projects finds agents fail mostly at discovering unstated requirements, and an open computer-use stack shows harness design moving scores as much as models.