Anthropic’s new Claude feature can leak data—users told to “monitor chats closely”

Summary

Independent AI researcher Simon Willison, reviewing the feature today on his blog, noted that Anthropic's advice to "monitor Claude while using the feature" amounts to "unfairly outsourcing the problem to Anthropic's users." Anthropic’s mitigations Anthropic is not completely ignoring the problem, however. The company has implemented several security measures for the file creation feature. For Pro and Max users, Anthropic disabled public sharing of conversations that use the file creation feature. For Enterprise users, the company implemented sandbox isolation so that environments are never shared between users. The company also limited task duration and container runtime "to avoid loops of malicious activity." For Team and Enterprise administrators, Anthropic also provides an allowlist of domains Claude can access, including api.anthropic.com, github.com, registry.npmjs.org, and pypi.org. The documentation states that "Claude can only be tricked into leaking data it has access to in a conversation via an individual user's prompt, project or activated connections." Anthropic's documentation states the company has "a continuous process for ongoing security testing and red-teaming of this feature." The company encourages organizations to "evaluate these protections against their specific security requirements when deciding whether to enable this feature." Prompt injections galore Even with Anthropic's security measures, Willison says he'll be cautious. "I plan to be cautious using this feature with any data that I very much don’t want to be leaked to a third party, if there’s even the slightest chance that a malicious instruction might sneak its way in," he wrote on his blog. We covered a similar potential prompt injection vulnerability with Anthropic's Claude for Chrome, which launched as a research preview last month. For enterprise customers considering Claude for sensitive business documents, Anthropic's decision to ship with documented vulnerabilities suggests co...