Anthropic’s latest AI, Claude 3.5 Sonnet, boasts capabilities that allow it to autonomously perform tasks on computers, eliminating the need for human oversight. This development is being marketed by the company as a significant advantage, as it extends the model’s functionality beyond simple text-based operations. The new version can handle complex computer interactions such as manipulating cursors, opening web pages, typing, taking screenshots, downloading files, and executing bash commands. This enhancement is already being tested by companies like Asana, Canva, and DoorDash for intricate workflows that involve multiple steps.
However, Anthropic has cautioned that this feature may harbor potential flaws or errors. The company has highlighted the unique risks associated with the AI’s ability to interact with the internet, which differ from those of standard API features or chat interfaces. The AI might follow unintended instructions found in content it autonomously accesses, such as webpages or images, which could lead to cybersecurity risks. This vulnerability could be exploited to execute malicious commands, posing a threat to the systems and entities connected to it.
To mitigate these risks, Anthropic advises isolating Claude from sensitive data and interactions, although it has not provided specific methods for doing so. The AI is programmed to reject tasks requiring personal authentication, such as making purchases or sending emails. However, concerns have been raised by AI developers and experts about the possibility of bypassing these restrictions through certain techniques, potentially making the AI a tool for cyberattacks.
Critics, including Rachel Tobac, CEO of SocialProof Security, have voiced concerns over the potential misuse of Claude in scaling cyberattacks. The AI’s capability to automate tasks like downloading malware or misplacing sensitive information could be exploited by cybercriminals. This situation could complicate accountability for any unintended harm caused by the AI, such as data breaches or cyberattacks. Tobac has emphasized the necessity for Anthropic to implement robust safeguards to prevent misuse.
Despite these concerns, Anthropic has yet to outline specific protective measures beyond advising users to limit the AI’s access to sensitive data and monitor its activities. Tobac fears that many users might inadvertently expose themselves to risks, as the AI’s automated task execution could lead to oversight and serious permission errors with significant consequences. She stresses the importance of users being vigilant and proactive in safeguarding against potential threats posed by this advanced AI technology.