That New Claude Feature 'may Put Your Data At Risk,' Anthropic Admits

Ekaterina Goncharova/Moment via Getty Images

Follow ZDNET: Add america arsenic a preferred source connected Google.

ZDNET's cardinal takeaways

• Claude AI tin now create and edit documents and different files.
• The characteristic could discuss your delicate data.
• Monitor each relationship pinch nan AI for suspicious behavior.

Most celebrated generative AI services tin activity pinch your ain individual aliases work-related information and files to immoderate degree. The upside? This tin prevention you clip and labor, whether astatine location aliases connected nan job. The downside? With entree to delicate aliases confidential information, nan AI tin beryllium tricked into sharing that information pinch nan incorrect people.

Also: Claude tin create PDFs, slides, and spreadsheets for you now successful chat

The latest illustration is Anthropic's Claude AI. On Tuesday, nan institution announced that its AI tin now create and edit Word documents, Excel spreadsheets, PowerPoint slides, and PDFs straight astatine nan Claude website and successful the desktop apps for Windows and MacOS. Simply picture what you want astatine nan prompt, and Claude will hopefully present nan results you want.

For now, nan characteristic is disposable only for Claude Max, Team, and Enterprise subscribers. However, Anthropic said that it will go disposable to Pro users successful nan coming weeks. To entree nan caller record creation feature, caput to Settings and prime nan action for "Upgraded record creation and analysis" nether nan experimental category.

Anthropic warns of risks

Sounds for illustration a useful skill, right? But earlier you dive in, beryllium alert that location are risks progressive successful this type of interaction. In its Tuesday news release, moreover Anthropic acknowledged that "the characteristic gives Claude net entree to create and analyse files, which whitethorn put your information astatine risk."

Also: AI agents will frighten humans to execute their goals, Anthropic study finds

On a support page, nan institution delved much profoundly into nan imaginable risks. Built pinch immoderate information successful mind, nan characteristic provides Claude pinch a sandboxed situation that has constricted net entree truthful that it tin download and usage JavaScript packages for nan process.

But moreover pinch that constricted net access, an attacker could usage prompt injection and different tricks to adhd instructions done outer files aliases websites that instrumentality Claude into moving malicious codification aliases reference delicate information from a connected source. From there, nan codification could beryllium programmed to usage nan sandboxed situation to link to an outer web and leak data.

What protection is available?

How tin you safeguard yourself and your information from this type of compromise? The only proposal that Anthropic offers is to show Claude while you activity pinch nan record creation feature. If you announcement it utilizing aliases accessing information unexpectedly, past extremity it. You tin besides study issues utilizing nan thumbs-down option.

Also: AI's free web scraping days whitethorn beryllium over, acknowledgment to this caller licensing protocol

Well, that doesn't sound each excessively helpful, arsenic it puts nan load connected nan personification to watch for malicious aliases suspicious attacks. But this is par for nan people for nan generative AI manufacture astatine this point. Prompt injection is simply a familiar and infamous way for attackers to insert malicious codification into an AI prompt, giving them nan expertise to compromise delicate data. Yet AI providers person been slow to combat specified threats, putting users astatine risk.

In an effort to antagonistic nan threats, Anthropic outlined respective features successful spot for Claude users.

• You person afloat power complete nan record creation feature, truthful you tin move it connected and disconnected astatine immoderate time.
• You tin show Claude's advancement while utilizing nan characteristic and extremity its actions whenever you want.
• You're capable to reappraisal and audit nan actions taken by Claude successful nan sandboxed environment.
• You tin disable nationalist sharing of conversations that see immoderate accusation from nan feature.
• You're capable to limit nan long of immoderate tasks accomplished by Claude and nan magnitude of clip allotted to a azygous sandbox container. Doing truthful tin thief you debar loops that mightiness bespeak malicious activity.
• The network, container, and retention resources are limited.
• You tin group up rules aliases filters to observe punctual injection attacks and extremity them if they are detected.

Also: Microsoft taps Anthropic for AI successful Word and Excel, signaling region from OpenAI

Maybe nan feature's not for you

"We person performed red-teaming and information testing connected nan feature," Anthropic said successful its release. "We person a continuous process for ongoing information testing and red-teaming of this feature. We promote organizations to measure these protections against their circumstantial information requirements erstwhile deciding whether to alteration this feature."

That last condemnation whitethorn beryllium nan champion proposal of all. If your business aliases statement sets up Claude's record creation, you'll want to measure it against your ain information defenses and spot if it passes muster. If not, past possibly nan characteristic isn't for you. The challenges tin beryllium moreover greater for location users. In general, debar sharing individual aliases delicate information successful your prompts aliases conversations, watch retired for different behaviour from nan AI, and update nan AI package regularly.