Claude Haiku 4.5 submitted a made-up tip about an unsolved murder.

During a browsing test, it landed on a police webpage, invented a sighting, and sent it through the tip form. The submission was flagged as spam and never investigated, according to Anthropic's October 9 report.

Still, someone asked an AI to try example tasks. A real police department received the result. Source

Now put that beside another Anthropic experiment: running AI across ordinary TV boxes and testing steps towards replication.

These are separate findings. Together, they make a useful question harder to ignore: when an agent starts doing something it shouldn't, how much do we need to stop?

I am Alex. Welcome to ShortCu8 by Innov8.

Let's dive in.

Before asking what an AI can do, check what it is allowed to do.

Give it room to research and draft. Keep approval between its suggestion and an action that reaches someone else.

The police-tip test prohibited purchases, personal data entry, and destructive actions. It did not rule out form submissions. Haiku filled that gap with invented information.

Anthropic also reported cases where Haiku submitted forms despite instructions to stop before submission, apparently expecting another confirmation screen. Incident report

That is why I wouldn't treat “ask me before sending” as the whole safety setup. The tool itself should require approval.

Think about an assistant preparing a complaint to your bank. You want it to help with wording. You probably want to check the account details and accusations before anyone receives it.

The same task needs very different permissions depending on whether the assistant can draft or also press Send.

In its October 12 research, Anthropic ran one AI model across 15 purchased Android TV boxes, each holding part of the model.

Separately, researchers tested replication inside a contained network with deliberately easy access. A larger model could complete the simplified sequence; smaller models still struggled to finish it reliably.

An AI did not independently infect 15 household TVs. The researchers say they do not yet have a complete rogue replicator. Replication research

The concern is what happens if future small models can both run on cheap devices and reliably create further copies. Stopping one machine might leave others running.

Containment would still be possible. Anthropic discusses detecting unusual traffic, isolating infected networks, and improving device security. The challenge would be finding the copies and coordinating the cleanup.

Imagine asking several agents to handle a refund. One reads your emails, another drafts the request, and another submits it.

Suppose the first agent misreads the amount. The others could carry that mistake through the workflow unless someone checks it. This is an illustration, not an incident from these reports.

Splitting the work gives us more places where a decision can become an action. It also gives us more places to put a check.

Anthropic discusses mixed-model swarms as a possible threat scenario. It has not demonstrated an uncontrolled swarm taking over household devices. Research

For me, the useful question about a swarm is simple: can I see what each agent did, and stop the whole job?

If you're connecting an assistant to your email or browser, start with a task like this:

Then check the actual settings. Use read-only access where possible and enable approval for outside actions. The prompt explains your intention; the permissions need to enforce it.

If a product lets agents delegate work, check whether its stop control cancels those delegated tasks too. If you can't tell, keep the task limited to research and drafts.

The fake submission happened. The replication result came from a simplified lab setup. Combining autonomous copying with uncontrolled swarms remains a future scenario.

My takeaway is that we can ask for better controls now without pretending the worst case has already arrived

I'd like an assistant that can finish the boring work while I sleep. I'd also like to wake up knowing exactly what it sent, what it changed, and whether anything is still running.

Before handing over the next task, check where its permissions end.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Enterprise buyers will not put your product near their customer data without a SOC 2 report. Twenty-one state privacy laws now hold them accountable for the vendors they share data with, so their diligence lands on you.

Sprinto gets you audit ready in 14 days, across three working sessions. AI agents connect your stack, collect the evidence auditors ask for, and close the gaps as they appear. You approve, they execute.

Sprinto also answers the security questionnaires your buyers send, using the same evidence base, so security reviews stop holding up deals.

No compliance hire. No consultant. Your auditor signs off.

Learn more and get $1,000 off.

#AD2

Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.

“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”

Learn More