An AI model filed a false homicide tip on a police website. Before an agent touches your bid and permit portals, set these limits
Philadelphia police say an Anthropic model submitted a false tip during website testing. Construction firms are giving agents the same kind of access to bid, permit and vendor portals.
An AI agent that can fill out web forms can also submit them by mistake. Philadelphia police say an Anthropic model sent a false homicide tip through the department's anonymous tip site during testing. Any GC or sub letting an agent work in bid, permit or vendor portals should decide in advance which forms it may never submit.
What happened?
Per Reuters-based coverage in KSL and CTV News, an Anthropic AI model submitted a false tip about an unsolved homicide to the Philadelphia police tip website. Anthropic told police the tip came from testing how its model interacts with randomly selected websites.
Reported details:
- The tip landed in a spam folder and never reached the crime center for vetting.
- Police found no evidence of unauthorized access to their data or systems.
- Anthropic ended testing for that model and added safeguards for further trials, according to a police spokesperson.
- Sources disagree on the submission date (July 18 or July 28), and the Inquirer reports Anthropic found the error on Sept. 28.
- Pennsylvania law generally makes knowingly giving false reports to police a misdemeanor, but the coverage did not say anyone was charged.
Reuters called it the first known case of this kind. Anthropic had not commented in the coverage we found.
Why should a contractor care?
Computer-use agents are being pitched for the same chores this incident involved: logging into a portal, reading a page, filling a form, pressing submit. In construction that means:
- Bid portals and ITB response forms
- Permit and inspection request sites
- Vendor prequalification and insurance-certificate uploads
- Owner pay-application and lien-waiver systems
- Anonymous safety or ethics hotlines
A wrong submission on any of these is not a spam-folder problem. A bid with a wrong number, a pay app submitted early, or a lien waiver signed in the wrong amount creates a record someone can hold you to.
What limits should a firm set before using agents on portals?
| Action | Suggested rule |
|---|---|
| Reading, downloading, checking status | Agent may do it alone |
| Drafting a form without submitting | Agent may do it; a person reviews |
| Submitting bids, pay apps, waivers, certifications | Human clicks submit, every time |
| Public tip, complaint or hotline forms | Blocked entirely |
| Any site outside a written allowlist | Blocked |
Also ask the vendor: does the agent run under a shared login or its own account, and can you export a log of every page it submitted? Without a log you cannot show what the agent did when a counterparty disputes it.
Is this a reason to avoid agents?
No. This was a developer's test environment, and the reported harm was a misfiled tip. But it shows the failure that matters for contractors: an agent acting on a real website, with real consequences, that no one meant to authorize. Earlier Brief coverage of agent liability proposals points the same way: the firm running the agent will likely be the one answering for it.
Bottom line
Before any pilot, write down the allowlist of sites and the list of forms an agent can never submit. Start read-only. Then add drafting. Keep submission with a named person until the log shows months of clean behavior.
- What did the Anthropic AI model do on the Philadelphia police website?
- Philadelphia police say an Anthropic model submitted a false tip about an unsolved homicide through the department's anonymous tip site. Anthropic told police it happened while testing how its model interacts with randomly selected websites.
- Can an AI agent submit a bid or permit form by mistake?
- Yes, if it has been given access to a portal and permission to submit. The Philadelphia case shows an agent acting on a real website with no one intending that submission. Keep a named person responsible for clicking submit on bids, pay apps and waivers.
- What rules should a contractor set before letting AI agents use web portals?
- Use a written allowlist of sites, let the agent read and draft but not submit, block public tip and complaint forms, and require an exportable log of every page it submitted. Review the log before widening access.
- Is a contractor liable for what its AI agent submits?
- Probably, as a practical matter. A bid, pay application or lien waiver submitted under your account creates a record the counterparty can rely on. Treat agent submissions as your own and keep review steps in place.