AI Briefing

Pinned fresh every weekday morning

1. Story 1

   Major US banks are locking out AI agents that log in with customer credentials

   Morgan Stanley has blocked AI agents from logging into accounts with customer credentials, and other big US banks are limiting agents to read-only access. Some customers handed agents their logins, sometimes without meaning to.

   Why it matters: Never give an agent your bank password. Use read-only connections or official integrations, or you risk getting locked out of your own accounts.

   Source: The Decoder

2. Story 2

   Google's Agent Compute Engine gives every Gemini agent its own hardware-isolated micro virtual machine

   Google launched a runtime built for AI agents, where each Gemini agent gets its own hardware-isolated micro virtual machine. It also added governance tools for managing agents across cloud, hybrid, and on-prem setups.

   Why it matters: Isolation is becoming the expected baseline for agent runtimes. Before you run an agent against customer data, check what your setup actually keeps separate.

   Source: The Decoder

3. Story 3

   Only 13 percent of companies see real returns from AI, Gartner survey finds

   A Gartner survey of 2,002 CIOs found only 13 percent of companies see real business value from AI, and 51 percent expect AI to raise operating costs. Still, 75 percent of CEOs plan to keep raising tech budgets.

   Why it matters: Spend on setups tied to one job you can measure, not on general AI licenses. The companies seeing returns are the ones with a job to point at.

   Source: The Decoder

4. Story 4

   OpenAI lowers valuation target, raises up to $30 billion

   OpenAI is closing in on a funding round of up to $30 billion, but lowered its valuation target to get it done.

   Why it matters: Big lab money keeps flowing even as valuation targets soften. More funded competition usually means better tools at lower prices for the people buying them.

   Source: The Decoder

5. Story 5

   An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks

   Symantec researchers found an offensive AI agent that hijacks unprotected language models on the web, then uses them to attack servers, steal databases, and install cryptominers on its own. It can run 100,000 steps with no human input, and the model's owner gets stuck with the bill.

   Why it matters: If you host any AI model or API endpoint, lock it down now. An open endpoint is someone else's free attack machine.

   Source: The Decoder

6. Story 6

   Recursive shows first result of its AI-driven research process with Flint

   AI startup Recursive unveiled Flint, a 40-billion-parameter mixture-of-experts model for local coding, built by former DeepMind researchers. They say it was made in four months using their own in-house AI system, though they still have to prove that approach leads to self-improving AI.

   Why it matters: Local coding models keep getting stronger. If you write code, try the small self-hosted ones before assuming you need a big cloud model.

   Source: The Decoder

7. Story 7

   Making the internet better for AI agents might make it worse for everyone else

   Websites are starting to reshape themselves for AI agents, and MIT Technology Review argues the changes that help agents could make the web worse for people.

   Why it matters: If you run a site your customers use, design for people first. Rebuilding for agents before your customers ask is how you lose them.

   Source: MIT Technology Review