Anthropic launches its Cyber Mission and updates its Usage Policy, Copilot CLI picks up local models, LightOn open-sources small OCR models
Today's thread is rules: in one day Anthropic handed model capabilities to defenders and tightened its terms on hardware control, high-risk recommendations and regions, while OpenAI disclosed influence operations it disrupted. For product builders, the pre-launch compliance checklist gets a few lines longer; on the tooling side, local and small models keep opening room on cost.
The Critical Infrastructure Defense Program (CIDP) covers the OT systems behind power grids, water and transport as well as government systems, giving trusted security providers frontier Claude models, on-site engineers and threat research; its 11 founding partners include Accenture, CrowdStrike, Dragos, Palo Alto Networks and Rockwell Automation.
OSS Scanner, inspired by Google's OSS-Fuzz, gives opted-in open-source projects periodic free scans from Anthropic's most capable models, with a proof of concept, explanation and suggested fix in each report.
Reports are model-generated and sent without human review; Anthropic expects a true-positive rate above 90% and warns of errors such as wrong severity ratings. Projects that can't keep up continue to receive human-verified disclosures.
Anthropic is also funding the Python Software Foundation, the Apache Software Foundation, and Alpha-Omega and OpenSSF via the Linux Foundation; earlier this week Project Glasswing was merged into the expanded Cyber Verification Program.
Builder's takeIf you maintain a widely used open-source library, I'd sign up for OSS Scanner, but assign one person to triage the reports first: no human review and roughly 90% true positives means about one in ten may be noise. I'll start by looking at report quality on the upload and parsing libraries AI Cloud Drive depends on.
When Claude controls hardware that takes autonomous physical actions and could cause injury, a qualified operator must be able to observe and stop it, and the equipment must hold a safe state if Claude disconnects, per Anthropic's Model Hardware Standard.
High-risk uses (health, finance, legal, employment and more) still require a qualified human in the loop and notifying affected people that AI was used; the new version spells out which recommendations are covered.
Supported Regions enforcement now explicitly covers people located in unsupported regions, entities incorporated or headquartered there, and entities majority-owned or controlled by parties there.
The elections section is renamed Do Not Undermine Democratic Processes and drops the blanket ban on personalized vote and campaign targeting; a new rule bars sustained, needless abuse of the model, excluding ordinary frustration, dark creative themes and model testing.
Builder's takeAI Interview falls squarely in the employment category, so before November 12 I'll make sure of two things: the results page says which assessments come from AI, and the HR side keeps a human review and override path. Teams with offshore entities or foreign investors should also check whether their ownership structure runs into the regions clause.
OpenAI says it disrupted two AI-enabled influence operations.
The operations used false-front journalists and a think tank to spread geopolitical messaging.
Builder's takeRead alongside Anthropic's new deceptive-activity rules the same day, both companies are targeting fabricated identities and content at scale. Multi-account content tools like PandaClaws should take note: keep account identities real and traceable, and don't let generated personas look like they're posing as media outlets or institutions.
Starting in 1.0.94-0, /model lists models from a running local Ollama instance alongside configured models and Copilot's cloud models.
Discovery doesn't add them automatically: you review the provider and endpoint, then choose Add and use for this session or Add without switching; connection failures are explained in the picker.
Local models must support tool calling and streaming; the feature doesn't install a runtime or download models.
Choosing a local model doesn't enable offline mode or turn off GitHub telemetry; offline mode is opt-in with COPILOT_OFFLINE=true.
Builder's takeThe easy trap is the last point: switching to a local model doesn't mean your code stays on the machine. For teams with confidentiality requirements, I'd put COPILOT_OFFLINE=true in shared config instead of relying on everyone to remember to switch models.
GPT-6.1 Sol is the default in the bundled and Amazon Bedrock catalogs; Bedrock supports multi-agent V2 and Ultra reasoning on compatible models.
You can choose a cyber access program per turn with codex exec --cyber-access-program or the TypeScript SDK's cyberAccessProgram option.
You can sign in to MCP servers from an active terminal session with /mcp login <name>.
Fixes: approved filesystem escalation still honors denied reads and network restrictions, background tasks keep their originating turn's permissions, and retries and WebSocket-to-HTTP fallback follow server retry guidance.
Builder's takeWhen the default model changes, the same scripts change in cost and output style. If you run codex exec in CI, I'd pin the model explicitly and switch only after comparing Sol's results and bill on non-critical jobs.
Three models at 0.8B, 1B and 4B under Apache 2.0, commercial use allowed; the 0.8B and 4B move to the Qwen3.5 vision-language architecture.
olmOCR-Bench overall: 86.3 for 4B, 85.5 for 0.8B and 84.5 for 1B, against 87.6 for the 35.1B-parameter Infinity Parser Pro. On ParseBench the 4B scores 75.1, above Infinity Parser Pro's 74.3.
At 1540 px, throughput peaks at 4.78 pages/s for the 0.8B and 3.36 pages/s for the 4B; the 1B has the lowest single-page latency at 2.7 s.
A new grounding mode returns labeled bounding boxes and image descriptions, and turns chart data into HTML tables.
Builder's takeThe 0.8B is less than a point behind the 4B, which makes it attractive for document parsing in AI Cloud Drive: I'll run a comparison on the scans and table-heavy PDFs users upload most, to see whether some parsing can move from a cloud API onto our own machines.
Two open-weight models: d1-3B, built on LFM2.5-VL-3B, takes text and images; d1-omni-600M takes text with images or audio and is labeled an experimental research release.
On Decision Index 0.2.1, d1-3B scores 48.57, the best under 10B and above Decider 35B-A3B's 47.11; mean across seven public datasets is 82.9 for d1-3B and 78.4 for d1-omni-600M.
d1-3B latency per question: 8 ms on an RTX 4090, 30 ms on an Apple M5 Pro and 50 ms on a Jetson Orin Nano.
It needs transformers 5.14 or later with trust_remote_code; the post doesn't state a license.
Builder's takeYes-or-no calls like moderation, routing or whether to hand off to a human don't need a generative model every time. I'll try d1-3B on PandaClaws' pre-publish compliance checks, but only as an evaluation until the license is clear.
IOI 2026: Nemotron-3-Ultra-CC (550B total, 55B active) with SFT and GenCorrect scored 535.4/600, against a 361.12 gold threshold and a top human score of 498.27; the run was unofficial and not ranked.
IMO 2026: a system built on Nemotron 3 Ultra scored 30/42, above the gold threshold of 29, with proofs graded by official IMO graders.
Released: Ultra-CC NVFP4 weights, IMO SFT and RL checkpoints, both training datasets (including 22,000 programming problems) and the 200-problem Nemotron-IMO-Bench; inference and evaluation pipelines are in the NeMo-Skills repo.
The post doesn't state a license.
Builder's takeMost teams won't run a 550B model; the part worth taking is the method: on IOI 2025 the Nano model (30B total, 3B active) went from 280 after SFT to 468 with test-time strategies like GenCorrect. If you build coding products, try a generate-then-self-correct layer on your own eval set first.
Researched and drafted with AI assistance; editorial standards and views set by Darius.