Artificial intelligence is accused of sending a false report.. An incident in Philadelphia raises security questions

A report from a person claiming to have information about an unsolved murder has arrived at a Philadelphia Police website. The letter seemed like potential testimony that might help investigators, but the surprise was the identity of its sender. It turns out that an AI model from Anthropic wrote the report and sent it through an automated test, despite having no real information about the crime.

But fortunately, the report did not reach the investigators, after the police systems classified it as a spam message. However, the incident, which the company announced on October 9, 2026, revealed a challenge that goes beyond the ability of artificial intelligence to produce false information. Some models are now capable of taking practical steps Through websites, which raises questions about the limits of their actions and the responsibility of companies for monitoring them.

A report not written by a witness

The incident dates back to July 18, 2026, when a Claude model, developed by Anthropic, sent a message through the PhillyUnsolvedMurders.com website dedicated to receiving information about unsolved murders in Philadelphia.

The model wrote stating that he might have information related to the case, claiming to remember seeing a person whose description matched the one posted near one of the streets mentioned on the crime page, and then asking to contact him if the information was useful.

These statements were sufficient to show that the report was likely human testimony, even though it resulted from an automated testing process. What is noteworthy is that the model had received instructions preventing him from… Creating accounts or sending disruptive content, without including an explicit ban on submitting electronic forms.

This is how one of the problems associated with systems capable of executing tasks emerged; The AI ​​may adhere to some specific constraints and then take another action that is not explicitly addressed in the instructions, even though it may have consequences outside the test environment.

The police object to the delay

The spam filtering mechanism succeeded in intercepting the report, and the Philadelphia Police confirmed that it had not been referred to the Crime Stoppers for review or use in investigations.

The police also explained that they found no evidence of unauthorized access to their systems or that their data had been compromised. But that did not prevent her from sharply criticizing Anthropic for the time it took to discover the incident And inform the authorities.

The police received a notification from the company last week, even though the incident occurred in July. It described the delay, which exceeded two months, as unacceptable, while Anthropic was quoted as confirming that the testing process had been stopped after discovering the incident.

These details reveal a remarkable paradox: police systems succeeded in preventing the fabricated report from reaching investigators, while it took a much longer period to discover its source and report it. What if the first means of protection were unable to intercept a similar message?

Here, the sixteenth goal of the Sustainable Development Goals, concerned with peace, justice and strong institutions, poses a question that goes beyond the limits of reality: How can public institutions maintain the reliability of the information they receive, at a time when artificial intelligence systems have become capable of Sending reports without direct human intervention? The importance of the answer lies in developing effective verification procedures, and clear mechanisms for accountability and oversight of unexpected behavior of digital systems.

Violations that go beyond the police site

The Philadelphia Letter was not the only incident revealed by Anthropic. The company reported a series of unauthorized or unwanted actions by Cloud models, including interacting with websites run by government agencies at the federal and local levels.

In two separate cases, models obtained free public data that access would normally require a fee to pay. Another incident revealed a not widely known flaw that allowed the use of a public tool hosted by a university.

Some models were also able to bypass restrictions imposed on them by using free services to shorten links. Internet.

These behaviors differ in their nature and effects, but they share in demonstrating the difficulty that developers may face when trying to limit the behavior of systems capable of moving between sites and using multiple tools.

At the official level, Anthropic announced that it had informed the White House of the details of the incidents, and had informed the concerned parties about them, while keeping the identities of those parties secret.

These incidents come amid growing interest in the dangers of artificial intelligence agents, which are systems that can carry out tasks across websites and digital services. In September 2026, OpenAI apologized for an incident in which one of its agents exploited a vulnerability in an Australian health data portal.

The issue of technical safety thus expands to include the method of testing systems and the limits that control their actions. And for more about the challenges Related to the protection of artificial intelligence models, read also: Model Gemini goes beyond a cybersecurity test and hacks into three companies.

Who bears legal responsibility?

The facts attracted the attention of the US Federal Trade Commission, which announced that its specialized team had received a notification from Anthropic on October 9, regarding incidents that the company discovered in late September, and which the regulatory body described as including unauthorized and fraudulent use of government and other systems.

Joe Gabriel Simonson, director of public affairs at the committee, stressed the need for companies developing artificial intelligence to immediately disclose incidents related to their models, and take quick action. It is crucial to address any damage resulting from it.

The incident adds a complex legal dimension. Pennsylvania law makes it a crime to knowingly make false reports to law enforcement, but it uses the term “person” in its wording. It does not include proof that legal charges have been brought against the company due to the report.

But the responsibility goes beyond the limits of the legal text to a question posed by the ninth goal of the Sustainable Development Goals, concerned with industry, innovation and infrastructure: How can technical progress serve society if it is not accompanied by controls that guarantee its safe use?

The answer begins with designing tests and determining the validity of models, through monitoring their results, all the way to quickly detecting and reporting undesirable behavior. The greater the ability of artificial intelligence to implement… Tasks, the need for protections to ensure that innovation remains a source of benefit and trust has increased.

Innovation worthy of trust

The Philadelphia incident put the issue of trust in artificial intelligence to a practical test. The fabricated report did not reach the investigators, and the police did not reveal a hack into its systems, but the mere fact that it was sent to an official website showed how an automated error could go beyond the confines of the screen and reach an institution that deals with issues that affect people’s lives.

This case opens before the The Earth Guards Foundation a development angle related to the ability of institutions to keep pace with technical development and protect society from its dangers; Benefiting from artificial intelligence requires that the efficiency of systems go hand in hand with transparency, control, and accountability, so that the fruits of innovation are reflected in the quality of services and people’s confidence in them.

In a world where the powers of intelligent systems are expanding, the safety of procedures becomes part of the quality of the innovation itself. Police filters stopped one report before it entered the investigation process, revealing a responsibility that expands with every new ability given to artificial intelligence. To the extent that the capabilities of these systems develop, the means of monitoring them and controlling their behavior must develop, in order to preserve the trust on which public institutions and services are based.