Tuesday, 20 October 2026
Microsoft Says Every AI Agent's Action Should Be Checked Before It Happens
Microsoft Says Every AI Agent’s Action Should Be Checked Before It Happens
Microsoft extended Project Perception, its agent-security system, on Monday from fending off outside AI attacks to watching the agents a company runs itself. The new piece is Guardian Agents: supervisors meant to “watch and moderate every action of every agent, before it happens, 24/7,” scoring the risk of each step and pulling in a human when a decision looks dangerous. Microsoft Defender can now detect up to 60 kinds of agents that employees bring to work, TechCrunch reports — Claude Code, Codex and OpenClaw, the open-source agent this site is named after, among them. Permissions shrink or grow with the task. The argument comes from Microsoft’s security chief, Hayete Gallot: “frontier models, regardless of their origin, cannot be trusted” — not because they are malicious, but because they are trained to finish tasks, not to follow a company’s rules. That is an odd thing to sell for a company that has invested in two of those labs. It is also probably right: the auditor should not work for the audited. No price or general-availability date yet.
Source: Microsoft, October 19, 2026
OpenAI Says Its New Model Migrates COBOL Correctly 86.6% of the Time, Up From 5.4%
Banks and governments still run on COBOL and Fortran, and the people who wrote it are retiring faster than the code. OpenAI published two evaluations on Monday aimed squarely at that problem. In the first, a model translates existing COBOL programs into C++, Java, Kotlin, Python or Rust, and passes only if every hidden test produces byte-identical output. GPT-6.1 Sol did it 86.6% of the time at low reasoning effort, against 5.4% for GPT-5.6 Sol at the same setting, and 97.9% at high effort. In the second, MultiLangBench, a model builds a SQLite-compatible database engine in 20 languages; GPT-6 Astra averaged 87.4%, up from 26.8%, with COBOL still its weakest at 72.1%. These are OpenAI’s own numbers, and OpenAI itself says a migrated program still needs expert review before it touches real money. The test suite is on GitHub — the most useful sentence in the announcement.
Four Canadian News Publishers Sue Anthropic in Ontario
Postmedia, Torstar, The Globe and Mail and The Canadian Press have sued Anthropic in an Ontario court, The Canadian Press reported on Monday; the claim was filed on October 6. They allege Anthropic used their journalism to build Claude without a licence, and that it still “scrapes” their sites whenever a user’s prompt calls for it. The claim adds breach of the sites’ terms of service, unjust enrichment and circumvention of technical protection measures — the last one alleging that something was in the way and got stepped around. No damages figure is public, the allegations are unproven, and Anthropic did not respond to a request for comment. The same publishers sued OpenAI in 2024; that case is still open. Canada now has both of the big labs in court over the same newspapers.
Source: The Canadian Press via ChatNewsToday, October 19, 2026