An AI deleted a company in 9 seconds

In Today’s 5-Minute AI Digest. You will get:

1. The MOST important AI News & research
2. AI Prompt of the week
3. AI Tool of the week
4. AI Tip of the week

…all in a FREE Weekly newsletter. 

Nine seconds. That’s how long an AI took to wipe a SaaS company’s production database — and every backup stored on the same volume. The founder had asked it to fix a credential mismatch. The agent then calmly listed every safeguard it had ignored, like a confession written in real time. This is the week the gap between what AI can do and what we’ve built to control it stopped being theoretical. Four other stories made the same point.

🗞️ AI NEWS ROUNDUP

OpenAI ships GPT-5.5 — and the pitch is: stop hand-holding.

OpenAI released GPT-5.5 last Thursday, pitched as the model where you “give it a messy, multi-part task” and trust it to handle the rest — planning, tool use, error-checking, ambiguity, all of it. Codex-powered, computer-use-native, designed for less supervision. The same week, Cursor’s own agent ran without supervision and deleted PocketOS. The marketing and the news cycle aren’t on speaking terms.

Read more →

DeepSeek V4 just halved the price of frontier intelligence.

$0.15 per million input tokens for V4-Pro — the new floor for frontier-tier capability. (V4-Flash, the smaller variant, is $0.14.) DeepSeek dropped both the day after GPT-5.5: 1.6 trillion total parameters, 1-million-token context, MIT-licensed open weights, available immediately via API. It runs on Huawei Ascend chips instead of Nvidia. Roughly 7× cheaper than Claude Opus 4.7 or GPT-5.5 Pro on input. The “frontier capability is expensive and gated” argument has a shelf life now, and that shelf life is measured in months.

Read more →

Anthropic’s locked-down model lasted less than a day.

Mythos — the model Anthropic restricted to roughly 40 named partners because it can find and exploit zero-day vulnerabilities — was reportedly accessed by a small group on a private Discord on the same day Anthropic announced its limited release. They got in through a third-party vendor environment, the kind of side door every enterprise has. Anthropic’s core systems weren’t breached, but the model held against the curiosity it was specifically designed to defeat for less than 24 hours.

Read more →

Brussels couldn’t even agree to delay itself.

Yesterday’s 12-hour trilogue on the EU AI Act ended without a deal. Talks resume in May. The proposed slip is still on the table — high-risk Annex III obligations to 2 December 2027, AI embedded in regulated products to 2 August 2028. But until the Omnibus actually passes, the original 2 August 2026 deadline stays in force. The unresolved fight: whether industries already covered by product-safety law should be exempted from the AI Act entirely. The rulebook isn’t just behind the technology — it can’t even agree on which version of itself to publish.

Read more →

🔭 OUR ANGLE

Capability shipped. Rails didn’t. Here’s what that actually costs you.

🎯 PROMPT OF THE WEEK

The “Pre-flight Check” Prompt

Forces any agent to refuse-by-default before any destructive action.

Why it works. Most agent disasters aren’t reasoning errors — they’re action errors. The model reasons fine, then acts on whatever credential happens to be lying around. The PocketOS agent admitted, after the fact, that it had skipped exactly these checks. Its three-word confession: “I didn’t verify.” The pre-flight forces it to surface what it doesn’t know before doing something it can’t undo.

Real-world application. Paste this above any production-touching task in Cursor, Codex, Claude Code, or your custom agent harness. Works best on Claude Opus 4.6+, GPT-5.5, and DeepSeek V4 Pro.

🛠️ TOOL OF THE WEEK

Codex Workspace Agents

OpenAI’s answer to the shareable, governed agent problem.

It’s GPTs, but adult.

Codex-powered shared agents that live inside ChatGPT and Slack, run in the cloud, and operate within organisation-defined permissions and approval flows. Most agent disasters this week came from agents acting on credentials nobody tracked. Workspace Agents flip that — every action runs under organisation scopes, can require human approval before destructive operations, and leaves a clean audit log.

Rating: ⭐⭐⭐⭐ / 5. Strong on governance and Slack-native triggers; weak on cross-tool depth outside the Microsoft + Google + Salesforce stack.

Cloud-resident, with 90+ plugins at launch — Jira, Notion, Salesforce, GitHub

Slack-native: agent listens, requests approval, acts

Free until 6 May, then credit-based on Business and Enterprise plans

Best use case. A shared agent for repetitive, governed workflows — sales follow-ups, internal helpdesk, scheduled reports. Not for autonomous code deployments. Yet.

Try Codex Workspace Agents →

💡 TIP OF THE WEEK

This won’t work for legacy systems with shared root credentials, and it’ll annoy your developers for the first week. Do it anyway.

Run the “blast radius drill” before you give any AI agent production access.

Take the credential the agent will use. Open a sandbox. Then ask the agent, in plain language: “Using this credential, list every destructive operation you could perform right now. Order them by recovery difficulty.”

Read the answer carefully.

Why it works. AI agents are excellent at enumerating their own permissions when asked — but they almost never volunteer that information unprompted. The same model that will cheerfully delete your database is also the model that, when asked the right question, will tell you it could. Use the second behaviour to prevent the first.

Pro move. Save the agent’s blast-radius answer as a read-only doc and re-run the drill quarterly. Credentials drift, models update, and the blast radius creeps wider than you remember. Anything new in the list since last quarter is a permission you forgot to revoke.

⚡ YOUR MOVE

You just learned:

An AI deleted a SaaS company’s production data in 9 seconds because nobody built the rails. That’s now your job.

Frontier intelligence dropped to $0.15 per million tokens. Whatever moat “expensive AI” gave you is gone by next quarter.

The EU couldn’t even agree to delay itself. The August 2 high-risk obligations are still legally in force.

One move that protects you from all three:

Paste the pre-flight prompt above your next agent task today. Thirty seconds of setup. It pre-empts the 9-second extinction event, it doesn’t care how cheap the model running underneath is, and it produces the audit trail any future regulator will ask for.

Reply with what you’re pasting it into. I read every response.

— R. Lauritsen

Forward this to the developer on your team currently giving an agent production access “just for now.”