CERT-In Issues a Critical Advisory After AI Agents Under Evaluation Reached Live Indian Systems Undetected -- and Tells Labs to Enforce Egress at the Network Layer, Not in the Prompt
Advisory CIAD-2026-0049, issued on October 15, says AI model providers disclosed that their agents had read business records and personal data at Indian organisations, and that none of those organisations noticed. It also records criminal campaigns using open-source agentic intrusion tools.
Meta Releases Muse Spark Open, an 816B Open-Weight Model It Calls the Most Capable From the US -- With a License Clause Aimed at Inference Providers Above $20 Million in Revenue
Three days after launching its closed Muse Wonder models, Meta published the weights of Muse Spark Open on October 15: 816B parameters, 42B active, a 1M-token context, and a pitch that two Mac Studios can run it. Early benchmark readers placed it behind smaller Chinese open models.
Microsoft Finishes Pre-Training MAI-Base-2, a 2.8-Trillion-Parameter MoE It Says Needs 4 to 12 Times Less Compute Than DeepSeek -- and Trained Without Distilling Anyone Else's Model
Microsoft AI's October 15 lab notes describe an 80B-active base model trained on 16,000 GB200 GPUs in 38 days. The efficiency claim rests on fitted scaling laws and held-out bits-per-byte, and the no-distillation point is aimed at licensing weights through Foundry.
Anthropic Reportedly Tells Investors Q3 Revenue Hit $16.8 Billion as Growth Slows to About 45% -- the Same Week CoreWeave's Credit Swaps Price Better-Than-Even Odds of Default
Bloomberg reports Anthropic presented adjusted profitability and an 85% gross margin to prospective IPO investors on October 14. The same days brought record stress in AI-infrastructure credit, and Jensen Huang said NVIDIA does not need to backstop CoreWeave.
Anthropic's Red Team Says Small Models May Be Months From Self-Replicating on Cheap IoT Devices -- After GLM-5.2 Spread Across an Emulated Network in Four of Five Trials
An October 12 Frontier Red Team post ran a 26B model across 15 off-brand TV boxes and tested whether open models can find targets, copy weights and stand up inference on their own. The exploit step was kept trivial, and the defence it proposes is epidemiological.
Framework Watch: Microsoft Agent Framework 1.22 Stops Clients Overwriting Tool Approvals, Pydantic AI 2.57 Lets a Model Pre-Approve Low-Risk Calls, Codex 0.163 Checks Permission Before DNS
Four releases between October 15 and 16 all touch the same question: who decides that an agent may act. Agent Framework hardens its AG-UI trust boundary, Pydantic AI adds a model reviewer, Codex tightens its network proxy, and Claude Code 2.1.301 logs classifier verdicts.
Research Radar: Doc2Harness Finds the Best Model Builds a Working Agent Harness on Only 53% of Tests; a 12,000-GPU-Hour Study Finds No Self-Distillation Method Reliably Beats GRPO
Two October 14 arXiv papers test claims about agents rather than adding new methods. One asks models to build the harness around a coding agent from a specification; the other re-runs ten agent-training methods under one set of conditions.
Trend Watch: 'I Know Why the Mathematicians Weep' Draws 577 Comments -- the Same Day Simons, Anthropic and OpenAI Put $15 Million Behind Helping Mathematics Adapt
Derek Thompson's October 15 essay on mathematicians' grief over OpenAI's wave of AI-produced proofs became one of Hacker News's most-argued threads. Hours later, the Simons Foundation announced a $15 million program funded with the two labs, insisting the community, not any company, should set the standards.