3 hours ago
4
Follow ZDNET: Add america arsenic a preferred source connected Google.
ZDNET's cardinal takeaways
- Claude AI tin present make and edit documents and different files.
- The diagnostic could compromise your delicate data.
- Monitor each enactment with the AI for suspicious behavior.
Most fashionable generative AI services tin enactment with your ain idiosyncratic oregon work-related information and files to immoderate degree. The upside? This tin prevention you clip and labor, whether astatine location oregon connected the job. The downside? With entree to delicate oregon confidential information, the AI tin beryllium tricked into sharing that information with the incorrect people.
Also: Claude tin make PDFs, slides, and spreadsheets for you present successful chat
The latest illustration is Anthropic's Claude AI. On Tuesday, the institution announced that its AI tin present make and edit Word documents, Excel spreadsheets, PowerPoint slides, and PDFs straight astatine the Claude website and successful the desktop apps for Windows and MacOS. Simply picture what you privation astatine the prompt, and Claude volition hopefully present the results you want.
For now, the diagnostic is disposable lone for Claude Max, Team, and Enterprise subscribers. However, Anthropic said that it volition go disposable to Pro users successful the coming weeks. To entree the caller record instauration feature, caput to Settings and prime the enactment for "Upgraded record instauration and analysis" nether the experimental category.
Anthropic warns of risks
Sounds similar a utile skill, right? But earlier you dive in, beryllium alert that determination are risks progressive successful this benignant of interaction. In its Tuesday quality release, adjacent Anthropic acknowledged that "the diagnostic gives Claude net entree to make and analyse files, which whitethorn enactment your information astatine risk."
Also: AI agents volition endanger humans to execute their goals, Anthropic study finds
On a enactment page, the institution delved much profoundly into the imaginable risks. Built with immoderate information successful mind, the diagnostic provides Claude with a sandboxed situation that has constricted net entree truthful that it tin download and usage JavaScript packages for the process.
But adjacent with that constricted net access, an attacker could usage prompt injection and different tricks to adhd instructions done outer files oregon websites that instrumentality Claude into moving malicious codification oregon speechmaking delicate information from a connected source. From there, the codification could beryllium programmed to usage the sandboxed situation to link to an outer web and leak data.
What extortion is available?
How tin you safeguard yourself and your information from this benignant of compromise? The lone proposal that Anthropic offers is to show Claude portion you enactment with the record instauration feature. If you announcement it utilizing oregon accessing information unexpectedly, past halt it. You tin besides study issues utilizing the thumbs-down option.
Well, that doesn't dependable each excessively helpful, arsenic it puts the load connected the idiosyncratic to ticker for malicious oregon suspicious attacks. But this is par for the people for the 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 the quality to compromise delicate data. Yet AI providers person been dilatory to combat specified threats, putting users astatine risk.
In an effort to antagonistic the threats, Anthropic outlined respective features successful spot for Claude users.
- You person afloat power implicit the record instauration feature, truthful you tin crook it connected and disconnected astatine immoderate time.
- You tin show Claude's advancement portion utilizing the diagnostic and halt its actions whenever you want.
- You're capable to reappraisal and audit the actions taken by Claude successful the sandboxed environment.
- You tin disable nationalist sharing of conversations that see immoderate accusation from the feature.
- You're capable to bounds the duration of immoderate tasks accomplished by Claude and the magnitude of clip allotted to a azygous sandbox container. Doing truthful tin assistance you debar loops that mightiness bespeak malicious activity.
- The network, container, and retention resources are limited.
- You tin acceptable up rules oregon filters to observe punctual injection attacks and halt them if they are detected.
Also: Microsoft taps Anthropic for AI successful Word and Excel, signaling region from OpenAI
Maybe the feature's not for you
"We person performed red-teaming and information investigating connected the feature," Anthropic said successful its release. "We person a continuous process for ongoing information investigating 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 the champion proposal of all. If your concern oregon enactment sets up Claude's record creation, you'll privation to measure it against your ain information defenses and spot if it passes muster. If not, past possibly the diagnostic isn't for you. The challenges tin beryllium adjacent greater for location users. In general, debar sharing idiosyncratic oregon delicate information successful your prompts oregon conversations, ticker retired for antithetic behaviour from the AI, and update the AI bundle regularly.
English (US) ·