AI progress — 20 October 2026

A hacker used a swarm of mostly autonomous AI agents to break into more than 100 websites, according to the security firm Norforge, with Reuters reporting that John Deere, SharkNinja and Cummins were among the victims. Norforge does not name the targets itself. If the findings hold, agent-driven intrusion at that scale is no longer hypothetical for defenders. Microsoft, separately, is telling enterprises to monitor employee-used agents as they would insider threats.

New models and papers

• Advancing coding in low-resource programming languages — OpenAI describes work on ChatGPT for less common programming languages, with new evaluations covering software engineering, SQLite and COBOL migration. Why it matters: teams maintaining legacy or niche codebases get evaluations aimed at their languages rather than only mainstream ones.
• DeepSeek-V4.1-Flash-1017 — an image-text-to-text model release from DeepSeek on Hugging Face (about 18,000 downloads so far). Why it matters: an open-weights multimodal option to evaluate; the listing alone does not establish how it compares with other models.
• Self-Evolving Coding Agents via Harness-Policy Optimization — a paper on improving coding agents by optimizing the harness around the model together with the policy model. Why it matters: it targets the harness, which the authors say jointly determines agent performance with the model, as a place to gain capability without a bigger model.

New tools and software

• Structured outputs for Anthropic Claude models on Amazon Bedrock — Bedrock now returns schema-compliant JSON for all Claude models via JSON Schema output format and strict tool use, on both the bedrock-runtime and bedrock-mantle endpoints. Why it matters: you can drop custom parsing and retry logic for malformed JSON in Bedrock-based pipelines.
• Copilot model policy API for enterprise, orgs, and teams — GET and PATCH REST endpoints to read and update Copilot model policies for enterprises, enterprise teams, enterprise apps and organizations. Why it matters: admins can manage which models are allowed in code and audit it, instead of clicking through settings.
• Claude Code v2.1.304 — adds a transcript line when a plugin skips one of a mod’s hooks, a decider: “person” option for tool.check answers, and an e.isReadOnly flag on tool.check events. Why it matters: mod authors can require a human approval that automation cannot grant, and can see when another plugin bypassed their hook.

Overall digest

• A hacker broke into John Deere and SharkNinja with an army of AI agents — Norforge reports more than 100 websites were breached by agents that mostly operated on their own. Why it matters: this is a vendor’s finding, but it is a concrete case for assuming attackers can parallelize intrusion attempts cheaply.
• Microsoft says AI agents are like insider threats — Microsoft wants enterprises to closely watch agents that employees use on systems not controlled by AI labs such as OpenAI and Anthropic. Why it matters: it frames agent governance as an access-control and monitoring problem for security teams choosing tooling.
• CoreWeave’s debt is becoming the whole AI industry’s problem — CoreWeave’s bonds slipped into distressed territory last week, with investors demanding 1,000 basis points more than government yields. Why it matters: anyone relying on capacity from debt-funded GPU clouds has reason to check counterparty risk.