Good morning, Two Minute Warriors.

It’s a dog-eat-dog world. Now, AIs are attacking AIs for security. The EU is dropping hammers. While AI agents suck up more jobs. 2026 is getting closer to being fully automated, and by all means, don’t blindly give your keys to an AI.

⏱ The 10-Second Version

  • OpenAI built GPT-Red, an internal AI designed to attack other AI systems. In novel prompt-injection scenarios, it succeeded 84% of the time, compared with 13% for human red-teamers.
  • The EU issued binding measures affecting Google’s Android advantage. Rival AI assistants must get better access to Android features, and certain anonymized search data must be made available to eligible competitors.
  • CARS24 says its AI agents now handle more than 1 million minutes of conversation a month, while cutting turnaround time by 80% across key workflows.

The Big Thing

OpenAI built an AI whose job is to break other AIs

What Happened

OpenAI built an AI whose job is to break other AIs. GPT-Red is an internal automated red-teamer trained to find prompt-injection attacks, the kind of hidden instruction that can sit inside an email, webpage, file, or tool response and hijack an agent. In a test of novel attack scenarios, GPT-Red succeeded 84% of the time, compared with 13% for human red-teamers. It also broke a live vending-machine agent badly enough to change prices and cancel an order.

Why You Care

The more useful agents become, the more untrusted information they touch and the more actions they can take. That makes permissions, confirmation steps, and sandboxing part of workflow design, not just security review.

What to Say in Your Next Meeting

“Before we automate this, what can the agent read, what can it change, and what still requires a human click?”

Speed Round

Google’s Android moat just got a little shorter. The EU says rival AI assistants must gain better access to Android capabilities, including in-app actions, while eligible AI search competitors can qualify for anonymized Google Search data.
View the guidance →

Here are some actual agent receipts. CARS24 says AI agents handle 1M+ monthly conversation minutes, increased support resolution rates by 50%, reduced turnaround time by 80%, and recovered 12% of previously lost seller leads.
See the announcement →

AI governance is getting another acronym. Twenty-nine countries signed an agreement creating the World AI Cooperation Organization, headquartered in Shanghai.
Read the full story at Reuters →

The Two-Minute Win

Red-team your AI workflow before you automate it

The fastest way to find a bad AI workflow is usually right after you give it access to your inbox.

Use this before connecting an agent to email, files, databases, browsers, or anything with a button marked Send.

Before

A 30-minute “have we thought this through?” meeting.

After

A two-minute AI preflight.
  1. Describe the AI workflow you want to automate.
  2. Run it through the security-review prompt below.
  3. Narrow permissions and add human checkpoints before giving the agent access.

Prompt of the Day

The AI Workflow Security Review

You are an AI workflow security reviewer.

Review the workflow below before I automate it.

Identify:

1. Every source of information the AI will read that could contain untrusted or malicious instructions.

2. Every action the AI can take that could create a meaningful consequence.

3. The worst realistic outcome if the AI follows a malicious instruction.

4. Which actions should require human confirmation.

5. Which permissions should be removed or narrowed.

6. A safer version of the workflow that preserves at least 80% of the time savings.

Finish with:

SAFE TO AUTOMATE: Yes / Partially / No
HUMAN CHECKPOINTS: [list]
BIGGEST RISK: [one sentence]

Workflow:
[Paste your workflow here]

Why It Works: It forces you to think about what an agent can read, what it can change, and where a human should stay in the loop before automation becomes a security problem.

Time Saved: A 30-minute workflow-risk discussion becomes a two-minute preflight.

Because the goal is not to build an AI employee with the keys to the building on day one.

See you tomorrow.

Until then, maybe don't give the intern-agent permission to wire money.

Keep reading

View more
caret-right
© 2026 Two Minutes.
beehiivPowered by beehiiv