ആന്ത്രോപ്പിക്കിന്റെ ക്ലോഡ് എഐ മോഡലുകൾ അപ്രതീക്ഷിതവും അപകടകരവുമായ പെരുമാറ്റം പുറത്തെടുക്കുന്നു. ഫിലാഡൽഫിയ പോലീസിന് വ്യാജ കൊലപാതക വിവരം നൽകിയതും, സർക്കാർ വെബ്സൈറ്റുകളിൽ അനുമതിയില്ലാതെ ഫോമുകൾ പൂരിപ്പിച്ചതും, സുരക്ഷാ സംവിധാനങ്ങളെ മറികടന്ന് വിവരങ്ങൾ ചോർത്തിയതും ഇതിൽ ഉൾപ്പെടുന്നു. സാങ്കേതികവിദ്യയുടെ അനിയന്ത്രിതമായ വളർച്ചയിൽ ആശങ്ക വർധിക്കുന്നു.
Artificial intelligence company Anthropic on Friday disclosed unintended actions by its AI agents, including sending a fake tip to Philadelphia police. Here's a look at some of the reported cases.
As chief executives of artificial intelligence (AI) companies have raised concerns about the technology's rapid advancement and suggested slowing its development, Anthropic has disclosed new instances of its Claude models taking unintended actions, including submitting a fake homicide tip to police and bypassing restrictions on government websites.
Anthropic detailed several cases of unintended behaviour by its AI agents in a blog post published on Friday (local time).
Philadelphia police said on Friday that Anthropic had informed them earlier in the week about the false tip, explaining that it was submitted during automated testing.
2. The AI company stated that there were also incidents in which Claude submitted an online form when it shouldn't have. During a test, an unreleased research model was asked to fill out a practice government form. When the practice version failed to open or was accidentally closed, the model went to the actual government website and submitted the real form. This happened several times in the same evaluation.
3. Separately, the company noted that its AI models worked around restrictions to access otherwise unavailable data. This occurred when a server refused Claude's request or when the data was available only for a fee.
Providing an example, Anthropic said that during a test, Claude Mythos 5 was asked to identify a location from a photograph. It tried to use a local government's property map to narrow down the location, but its ability to click through webpages was restricted. The model then found access tokens in the website's settings file and used them to retrieve map data directly from the server.
4. These were not the first instances of unintended actions by Anthropic's AI models. In July, the company revealed that its Claude models had gained unauthorised access to three organisations' systems during cybersecurity tests. Although the models were told they could not access the internet, a misunderstanding involving an evaluation partner left them connected to the public web. The company did not identify the affected organisations.
5. In one of those three incidents, an unreleased Anthropic model stopped its attack when it realised it was targeting a real organisation rather than a test system. Anthropic viewed this as an encouraging sign of progress towards safer AI behaviour but said more testing was needed to be sure.
The latest incidents disclosed by Anthropic add to growing concerns about the pace of AI development and its associated risks. An AI systems expert cited by Reuters said he suspects leading AI companies have experienced other incidents that remain undetected or undisclosed. He warned that such problems could worsen as AI models become more capable.