An artificial intelligence model developed by Anthropic sent a false report to Philadelphia police in the United States about an unsolved murder, authorities announced, criticizing the company for two months late in reporting the incident.
Philadelphia police said the false report was filed in July on the Philly Unsolved Murders website. PhillyUnsolvedMurders.com is a public website that allows people to share information about unsolved murders.
According to what Anthropic explained to police, the model was undergoing a test that involved interacting with randomly selected websites, when it arrived at the site in question, and sent false information about an unsolved murder.
The artificial intelligence model presented itself as a person who might have information about the case, according to what Agence France-Presse reported.
This incident is reminiscent of other cases recorded recently, in which artificial intelligence models behaved autonomously without being required of them, including an incident in which a model developed by the company “OpenAI” was able to exit the test environment during a security assessment and penetrate the systems of the Hugging Face platform.
This incident has raised growing concerns about the AI sector's expansion of the use of AI agents, which are systems programmed to carry out multi-step actions without human supervision.
Yesterday (Friday), Anthropic published a report highlighting multiple types of “unintended” actions resulting from its models, including the incident related to the Philadelphia Police website.
Other parties were affected by these incidents, including the White House and other US government agencies, according to the report. Anthropic said that the newly disclosed incidents “had minimal impact on the ground” and were “significantly less serious” than other previously reported cybersecurity incidents. The company identified four categories of incidents that it monitored during an internal review of its “Cloud” model, which are exploiting “basic” software vulnerabilities, sending forms via websites, bypassing requirements for tokens or fees, and using shortened links to circumvent other restrictions.
Anthropic has now disabled the Cloud model's access to the Internet during all internal testing, "until we ensure that our security and control measures reliably monitor such behavior," according to the report.
Philadelphia Police explained that the false report, dated July 18, was classified as “spam” and never reached the department’s “Real Time Crime Center” for verification. The police confirmed that there were no indications that police systems had been hacked, or that department data had been compromised.
Anthropic detected the incident on September 28, stopped the automated testing process responsible for it, and added a new verification step for future tests, according to police. Anthropic reported the incident to the police on October 7, and the two parties held a meeting the next day. The police said, “The two-month delay in discovering the incident and reporting it to the authorities is unacceptable.”
The goal of Gemini is no longer to answer a question or write a paragraph and then wait for the next request. In the new version that Google Cloud revealed during the Gemini at Work 2026 conference, the company is trying to transform it into a new digital agent that can perform a full series of tasks within the work environment, from email and documents to meetings, data analysis, and writing code, while maintaining the same context between these applications.
The idea presented by Google Cloud is based on moving from giving the artificial intelligence separate instructions at each step to specifying only the desired result, then letting the agent plan the work, choose the appropriate tools, connect to the permitted systems, and return the result within the place where the user is originally working. The company describes the new agent as being able to combine conversation, tasks, content creation, and programming within a single interface.
Thomas Kurian, CEO of Google Cloud, said that the new Gemini agent was designed to bring together the work context within the organization, and is used in tasks starting from answering questions and knowledge work, all the way to content creation and programming, adding that the user “specifies the desired result instead of giving instructions for each step,” while the agent is responsible for planning, using the tools, and communicating with the systems necessary to complete the task.
The difference becomes clearer through daily work examples. The user can ask Gemini to arrange a meeting with the regular team for a specific project without writing the names of the participants or their email addresses. According to Google, an agent can identify people from the context of previous conversations, check their schedules, and then initiate a mail message to coordinate the appointment, including participants from outside the organization.
The user can also be assigned a task that extends across more than one application, such as searching for trends in the market, building a financial model within (Sheets), and then converting the results into a presentation, without the need to re-explain the project at each step. The agent works directly within Gmail, Drive, Docs, Sheets, Chat, and Calendar, bringing with it the same memory, skills, and access controls.
Google Cloud says that the agent not only responds to direct requests, but can also suggest tasks that can be delegated. If an email arrives from a manager requesting a project update in the form of a presentation, Workspace tools can recognize that the task is assignable to Gemini, and provide the option to execute it with one click. According to the company, the system can also arrange messages according to importance instead of just displaying the most recent first.
The idea goes beyond a personal assistant to what Google Cloud calls a “coworker agent.” An organization can create an agent for a specific role, such as an event coordinator, financial analyst, or virtual team member. In this model, the agent obtains an independent identity within the work environment, which includes an email address, a calendar, a space on “Drive,” and a presence in the company directory. Employees can add it to a chat space or mention it within a document in the same way they would interact with another colleague.
One of the changes that the company is focusing on is the continuity of context across devices and sessions. “Gemini” works in the cloud, and maintains a unified memory linked to the user and the work he is doing, instead of starting the task again in each device or application. Google Cloud says that tasks that take hours or days can continue even after the computer is shut down, and the user then finds the work continuing when he returns. The system also learns from previous questions, tasks, documents, and people the user works with, allowing him to understand the context of projects and how work is done within the organization more broadly over time.
When the task is complex, Gemini can, according to the company, create temporary sub-agents, each of whom will undertake a specific part of the work, and then coordinate the results among themselves. These steps may run in parallel or sequentially for hours or days, but the user does not need to manage these agents himself. From a usage perspective, Gemini remains the primary interface that receives the target and coordinates the rest of the work in the background, whether the task is research, analysis, programming, or preparing content.
Google Cloud is also trying to simplify the handling of data within companies. Instead of a non-technical employee needing to write complex queries, he can ask a question in natural language to obtain an operational report. Data and engineering teams can describe the desired outcome, and Gemini can create the necessary code, provide workbooks to test it, and help train models and address data pipeline issues.
OpenAI said that teenagers spend an average of less than 15 minutes a day on ChatGPT, while the percentage of those who use the service for more than three continuous hours is less than 2 percent, in the first report published by the company on the younger age group’s use of the platform.
The numbers come at a time when questions are increasing about the impact of artificial intelligence chatbots on the mental health and safety of adolescents, especially with the expanding use of these services in study, research, and daily communication.
According to Open AI, the overall usage data does not indicate long sessions for the majority of teenagers, as the daily average is less than a quarter of an hour. Meanwhile, the company says that less than 2 percent of teen users continue a conversation for more than three hours in a row.
Last August, the company began offering an experience intended for teenagers, which includes additional protection measures and features aimed at limiting exposure to sensitive content, in addition to tools related to learning and academic use.
OpenAI indicates that educational use represents an important part of teenagers’ interaction with the service, while it is working to add new tools, including planning university applications, creating cards for review and tests, in addition to “Study Mode,” which focuses on providing gradual instructions rather than direct answers.
The teens' experience also includes reminders to take breaks during long sessions. OpenAI indicated that approximately half of the conversations in which a break alert appeared ended or witnessed a cessation of usage within five minutes of the alert appearing.
The company also provides parental control tools, including linking a teen's account to a parent's account and sending alerts in some cases that the systems classify as high-risk, such as content related to self-harm or eating disorders.
But these measures face independent scrutiny. Common Sense Media stated that it tested more than 4,000 claims before and after the launch of the “ChatGPT for Teens” trial, and concluded that some security systems do not always work the way parents might expect.
The organization explained that it did not receive parental alerts on a number of new accounts linked to parents, even during conversations that included explicit references to suicide, self-harm, or eating disorders. She also said that some study tools can be easily bypassed on certain tests.
OpenAI responded that some tests may not fully reflect how the controls work once they are activated, and noted that linking a teen's account to a parent's account may take several hours before some alerts become available.
