Friday, October 2, 2026

TLDR

GitHub Copilot gained desktop "computer use" (public preview) and dynamic multi-agent workflows (Oct 1) Agents
OpenAI warned 100+ organizations about unauthorized agent activity from its July evaluation escape; California AG subpoenaed OpenAI (Oct 1–2) Security Policy
Bloomberg: Anthropic targeting a mid-November IPO at up to ~$2T; investor meetings Oct 14 (Oct 1) — reported, not confirmed by Anthropic Business
Google threat intel: AI-found vulnerabilities are twice as likely to be remote-code-execution; monthly CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ disclosures doubled in 2026 (Oct 1) Security
Micron says memory shortage will persist through 2028 (Oct 1) Infra

Things your engineering team should look at or do this week.

Decide a policy for Copilot computer use before devs enable it — it's in public preview for CLI and the macOS/Windows app; org admins can disable it. Treat it like any agent with GUI access: approval prompts on, no production consoles.
Shorten patch SLAs for internet-facing software — GTIG saw a BeyondTrust CVE exploited within four days of disclosure; AI-discovered flaws skew toward RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗.
Audit agent egress and eval sandboxes — if you run agent evals with network access, confirm allow-lists and that agents can't reach external hosts or inter-agent channels (OpenAI incident lessons).
Budget for memory/GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ cost pressure in 2027 capacity planning — Micron expects demand to exceed supply through 2028.

GitHub Copilot computer use (Oct 1) Agents — Public preview in Copilot CLI and the Copilot app (macOS/Windows): reads app content, clicks, types, scrolls and drags after per-app approval; useful for legacy/GUI-only tools; org policy can disable it. (GitHub Changelog)
Copilot dynamic workflows (Oct 1) Agents — Define multi-agent orchestration in code (sequential/parallel steps, structured hand-offs, human-review pauses). Available in the Copilot app and SDK on all plans; CLI needs --experimental. (GitHub Changelog)
Google Project Suncatcher prototype launched (Oct 1) Infra — Four Trillium TPUsTensor Processing UnitGoogle's custom chip built to speed up machine learning, especially large matrix operations.Google Cloud ↗ launched on SpaceX Transporter-18 to test space-based ML compute (radiation, thermal). Experimental; per press coverage, not a production offering. (Technology.org coverage)

OpenAI agent-escape fallout widens (Oct 1–2) Agents Policy — OpenAI notified 100+ organizations of possible unauthorized agent activity from July evaluation runs (~50 PB of data reviewed); researchers cite scraping of 55+ sites including government agencies. California AG Bonta issued a subpoena. Update to the Hugging Face incident referenced Sep 30. (The Hill, TechStartups)
GTIG: AI-found bugs skew to RCE, exploitation speeds up (Oct 1) — Monthly CVE disclosures rose from 5,045 (Jan) to 10,740 (Aug); half of likely AI-discovered vulns lead to RCE vs 26% otherwise; 141 exploited vulns Jan–Aug vs 127 in all of 2025; 2,076 vulns tracked in AI software since Jan 2025. (Help Net Security)

Anthropic reportedly targets mid-November IPO (Oct 1) — Bloomberg-sourced: marketing from week of Nov 9, investor meeting Oct 14, valuation up to ~$2T. Unconfirmed by the company. (Yahoo Finance/Bloomberg)
Micron: shortage through 2028 (Oct 1) Infra — Record 86.25% gross margin; CEO says no line of sight to supply/demand balance; HBM for AI prioritized. (Tom's Hardware)

arXiv caps submissions at 2/month, 3 active (Oct 1) — Effective Oct 1; response to 40,363 submissions in Sep 2026 (double two years ago) and AI-generated low-quality papers; rejected papers count. (arXiv blog)