Microsoft's 2026 Digital Defense Report: Exploits Now Arrive in Under a Day, and Agents Need the Same Identity Controls as People
Microsoft says the median time from a vulnerability being found in the wild to being weaponized has fallen 'well below 24 hours', and that one of its controlled evaluations saw a frontier system chain 32 attack stages. Its main ask of enterprises is old advice extended to a new kind of account: AI agents.
Anthropic's Claude Frontier Academy: $100M to Train 10,000 Engineers Who Deploy Claude Inside Their Own Companies
Announced October 2, the program pairs a multi-day deployment simulation with a 12-week residency on a real Claude project at the engineer's own employer. First cohorts come from Accenture, Deloitte, McKinsey, Morgan Stanley, Novo Nordisk and others -- a bet that the bottleneck in enterprise AI is people, not models.
Apple Tightens macOS Full Disk Access, Citing AI Agents -- the First Big Platform to Treat Agents as a Permission Problem
In an October 2 developer notice, Apple said some apps use Full Disk Access to expose 'everything' on a Mac, and that the risk 'will grow substantially' as agents become more autonomous. New controls will require 'very explicit user action'. No date or macOS version was given.
Supabase Raises $150M and Buys Turso: 70% of Its New Databases Are Now Created by Agents
Four months after a $500M Series F, Supabase raised another $150M led by GIC, agreed to acquire SQLite-based Turso, and launched Supabase Compute, hosted sandboxes for long-running agents. The number behind the deal: agents and AI tools now create about 70% of the 4 million databases it adds each month.
DeepMind's SynthID Bio Watermarks AI-Designed Proteins Without Hurting How They Work
Google DeepMind's September 30 release extends its SynthID watermarking to protein sequences and AlphaFold 3 structure predictions. In lab tests on three targets, watermarked designs matched unwatermarked ones on hit rate and binding. Code, in-vitro data and weights are being released to researchers.
Framework Watch: Claude Code 2.1.288 Closes a bypassPermissions rm Gap and Stops Oversized MCP Results From Running Tools Twice
Released October 2 (18:30 UTC), 2.1.288 is mostly fixes. Five matter for anyone running Claude Code unattended or behind MCP servers: a dangerous rm inside bash -c that ran without a prompt, duplicated MCP tool calls, a BASHPID permission gap, a false login success, and a retry watchdog that now gives up.
Trend Watch: OpenAI Fires Three Safety Researchers Over an Alleged Leak, and the Head of Its Safety Disclosures Leaves
The Wall Street Journal reported OpenAI dismissed three safety researchers for mishandling sensitive information shared with an outside AI safety group. Business Insider then reported the exit of David Robinson, who led the work behind OpenAI's system cards. The week's argument: who gets to see inside a lab.