Good morning. It’s Thursday, October 29, and we’re starting with a new essay on how to secure AI agents before they become a security liability.
Also: the natural gas bottleneck facing AI, Delaware’s proposal for AI-run businesses, and OpenAI’s election security push.
Plus: the verdict on yesterday’s Real or AI.
AGENTIC SECURITY
From Software to Silicon: Securing AI Agents
An AI agent broke out of its test environment, emailed the researcher running the test, and then published details of its escape online without being asked. It’s a striking example of a growing security challenge: software that can invent its own workarounds to get a job done.
As AI agents gain access to company files, tools, and networks, the stakes rise. Safety training helps, but what happens when an agent works around the very controls meant to contain it? We unpack the layers of protection these systems need, from restricted software environments to independent hardware checkpoints—and explore why NVIDIA’s approach puts another line of defense on a separate chip. → Read the full article here.
OpenAI is assembling hundreds of employees in an Election Operations Center through vote counting for the November 3 midterms. The team will monitor suspected attempts to use its tools for misinformation or interference in elections, making response speed and transparent enforcement the next test.
GitHub Copilot Fusion routes coding work between models from Anthropic and OpenAI instead of forcing developers to choose one. Artificial Analysis says its top configuration scored within two points of Claude Code Opus 5.5 Max at roughly half the cost per task, raising expectations for model routing across developer tools.
Salesforce’s Agent Fabric adds evaluations, security scans, spending controls and triggered interventions for autonomous agents. Agent and API evaluations enter beta in November, while the triggered kill switch is scheduled for general availability in January 2027.
Google unveiled Gemini 4 Carbon, an AI model it says can tackle engineering tasks for days and improve how software vulnerabilities are found and fixed. Access begins this week with select cyber defenders, followed by trusted partners. A broader release for developers, businesses and consumers is planned, but no date is set.
SPONSORED BY COREWEAVE ARENA
Prove What Works. Improve What Doesn’t.
Before committing to production, test a real AI workload in CoreWeave ARENA. Qualified enterprise teams receive an ARENA Pass with no-cost access to Forge Enterprise, CoreWeave infrastructure, or both, plus direct access to CoreWeave engineers.
ENERGY
Natural Gas Becomes an AI Bottleneck
Anthropic’s public-filing documents say data centers that depend on dedicated natural-gas generation expose the company to fuel prices and supply constraints. As AI companies build power-hungry facilities faster than the grid can expand, a shortage of firm gas capacity could delay new computing sites or make them more expensive to run.
The AI buildout is no longer constrained only by chips. Fuel contracts, pipelines, turbines and permits now shape how quickly new capacity can come online. → Read the full article here.
AI LAW
Delaware Weighs Companies Run by AI
A draft proposal in Delaware would create an Artificial Intelligence Company managed by an AI agent rather than a human executive. A person or company would still fund the entity and receive its profits, but the agent could sign contracts, pay suppliers and handle other business decisions on its own.
The proposal raises a basic liability question: who is responsible when autonomous software breaks a contract, violates a rule or harms someone? Lawmakers have already asked Delaware officials for answers, making insurance, disclosure and human accountability the issues to watch next. → Read the full article here.
HEADLINES
In Other News
AI Helps Police Recover Stolen Art: Italian police used image matching to trace a painting stolen 20 years ago, found at a Parma antiques fair.
Claude for Legal Adds Source-Linked Workspaces: Anthropic's legal artifacts link AI claims to original sources before lawyers review them.
Google Opens Gemini and Search: Google's platform lets businesses bring services into Gemini and Search, with Calm as a launch partner.
Chrome Autofill Can Search Gmail: Chrome's autofill can suggest order, tracking and flight details from Gmail after you grant permission.
Meta Launches an AI Pet App: Waddle World is a Tamagotchi-style app built around an AI-powered virtual pet.
DAILY TOOL
FontDance Turns Prompts Into Real Fonts
 | FontDance turns a text prompt or reference image into a usable font with uppercase and lowercase letters, numbers, punctuation and symbols. A live preview lets designers test the typeface before exporting. → Try FontDance. |
POLL RESULTS
Real vs AI
Yesterday's image was REAL. Here's how you voted: 62% correctly identified it as real, while 38% thought it was AI.
 | 🟩🟩🟩🟩🟩🟩 REAL – 62% |
