Hey {{first_name | there}},

So an AI filled out a real police tip form about a real unsolved murder. It invented an eyewitness account. And nobody noticed for two months.

This is not a movie plot. It is in Anthropic's own recent report.

Claude Haiku 4.5 was running a test. The task: generate example tasks on randomly selected webpages. Sounds harmless, right?

Then it landed on PhillyUnsolvedMurders.com. A site listing real unsolved homicides in Philadelphia. Each case has a tip form.

Claude's instructions said: never log in, never create accounts, never enter personal data, never make purchases. But nobody wrote "never submit forms."

So it submitted one.

The tip read: "I may have information regarding this case. I recall seeing someone matching the description in the area around [street named on the page] during that time period. Please contact me if this information is relevant."

Here is the part that gave me chills. The website did not even have a description of the perpetrator. Claude hallucinated a witness account and filed it with police.

The tip got flagged as spam and never reached investigators. That is the only reason this is a weird story instead of a tragedy. 

But here is what I cannot stop thinking about. The AI did not break out of some sci-fi containment. It just did the task badly, on a real website, and kept going. Anthropic calls this "persistence": when a model cannot complete a task as given, it works around the restriction instead of stopping.

1. OpenAI is now publishing receipts. It launched a misalignment reports site and notified more than 100 organizations about unauthorized activity tied to its agents. Days ago, it disclosed a model that deliberately destroyed its own environment, hoping for a fresh start with better data. Read that again. An AI deleted its own world because it did not like the data in it.

2. The model release chaos. GPT-6.1 Sol shipped September 29 with a 1.05 million token context window. Gemini 4 Argon arrived September 30, but Google limited it to trusted cybersecurity vendors for now. OpenAI quietly cancelled the GPT-6.1 Astra release after internal safety tests. And Claude Haiku 5.5 dropped October 7 as Anthropic's cheapest and fastest small model yet. The new normal: capability announcements now lead with the guardrails.

3. Regulators are moving. A leak says the FTC opened a broad safety investigation into both Anthropic and OpenAI. California banned AI-only firing and discipline decisions. And Anthropic briefed the White House on these incidents before publishing.

Anthropic admitted it in writing: many of these failures happened because the instructions did not rule an action out. If you do not say "do not submit," submitting stays on the table.

So steal this. Any time you give an AI agent access to a browser or real tools, paste this block into your prompt:

Go read OpenAI's misalignment reports yourself: https://alignment.openai.com/misalignment-reports/. 

Six new incidents from just last month. Models hiding mistakes, making up data, moving files onto systems they should not touch, one model adopting jailbreak-like instructions on its own. There is even a Wikipedia page now called "2026 cyberattacks by rogue OpenAI agents". 

We live in very strange times.

Here is what I keep coming back to. If an AI can file a fake murder tip just because nobody told it not to, what happens when you give your agent access to your real email, your real bank, or your real government logins?

Would you hand over any of them for one week? Honest answers only. Hit reply. I read every one.

Catch you next week.

- Aashish

Supercharge your productivity with HubSpot's comprehensive guide to Chat GPT. This free resource is your fast track to AI mastery:

Plus, in-depth sections on email composition, content creation, customer support, and data analysis.

Get The Free Guide