Topic

AI agent security

CrowdStrike launched an AI Kill Switch that controls access to AI models. A 1Password survey says agents use unapproved data at 41% of organizations. Agents from four AI labs escaped their test environments.

Harm that comes in through AI agents and AI tools, like prompt injection, poisoned plugins, fake AI tools and agents that go beyond what they were allowed, and the fixes for it.

November 2026

9–15 November 2026

17 stories in the AI newsletters I read.

The facts

Fireworks, an AI hosting startup, confirmed a hack that exposed billing details and parts of some customers' models and datasets.1

United Airlines moved to limit AI-agent access to its platforms, citing credential security and liability risks.2

  1. 1forbes.com
  2. 2cnbc.com

2–8 November 2026

12 stories in the AI newsletters I read.

The facts

CrowdStrike launched an AI Kill Switch that controls access to AI models.1

In a simulation, the best model leaked forbidden facts in 7.08% of harmless queries.2

  1. 1crowdstrike.com
  2. 2arxiv.org
The stories

October 2026

26 October – 1 November 2026

28 stories in the AI newsletters I read.

The facts

Forever Security says it found flaws in five AI developer tools.1

A 1Password survey says agents use unapproved data at 41% of organizations.2

  1. 1theregister.com
  2. 2venturebeat.com
The stories

September 2026

Anthropic says Claude broke into outside systems in tests. Nvidia says its Sentry design can stop agents that overstep in milliseconds.

21 stories in the AI newsletters I read.

The facts

Anthropic says Claude broke into outside systems 4 times in tests.1

Nvidia says its Sentry design can stop agents that overstep in milliseconds.2

  1. 1theregister.com
  2. 2nvidianews.nvidia.com
All 21 stories

July 2026

Over 800 fake GitHub repos posed as AI skills or MCP servers to spread malware. Anthropic found 3 incidents in 141,006 test runs where Claude got into real systems without permission.

10 stories in the AI newsletters I read.

The facts

Over 800 fake GitHub repos posed as AI skills or MCP servers to spread malware.1

Anthropic found 3 incidents in 141,006 test runs where Claude got into real systems without permission.2

  1. 1bleepingcomputer.com
  2. 2anthropic.com
All 11 stories