OpenAI announced a new feature that it says will provide additional protection from prompt injection attacks, where malicious chatbot instructions are hidden in webpages and other content sources.

Among other things, Lockdown Mode will disable live web browsing (so you can only access cached content), the retrieval and display of images from the web (you can still generate images), deep research, and agent mode.

The company says that even with Lockdown Mode turned on, ChatGPT could still be vulnerable to prompt injections — which could, for example, “appear in cached web content or in an uploaded file, and could still affect the behavior or accuracy of a response.”

But the goal is to reduce the likelihood that sensitive data gets shared in the process.

“Lockdown Mode is not intended for everyone,” OpenAI says. “It is designed for people and organizations that handle sensitive data and want stricter protection from data exfiltration risks related to prompt injection.”

The company says it’s currently rolling Lockdown Mode out to self-serve ChatGPT Business accounts, as well as eligible personal accounts.

Google announced on Tuesday that Android is launching fake call detection to protect against AI deepfake impersonation scams. The feature is rolling out globally in Phone by Google to Android 12+ devices this month, starting with Pixel devices.

As people increasingly refuse to answer calls from unknown numbers, scammers are shifting their tactics by spoofing trusted phone numbers and using AI deepfake technology to sound like authority figures, family members, or employers.

For example, a person may receive a phone call showing the caller ID “Mom,” and the voice may sound exactly like her, but the caller is actually a scammer using AI tools to impersonate her and request money for a fake emergency.

The new feature is on by default and works automatically behind the scenes.

Google explains that the new feature works kind of like a “digital handshake between devices.” When a contact calls you, and you’re both using Phone by Google, their phone sends a silent confirmation signal to your device to verify the call is legitimate and actually coming from their phone.

“If a scammer tries to impersonate your trusted contact, that initial confirmation signal will be missing,” Google explained in a blog post. “Your device will instantly notice this and ping your contact’s actual device to double-check. If their real device says, ‘I’m not making a call right now,’ you’ll get a warning on your screen advising you to hang up immediately.”

The tech giant notes that it built this feature on top of Rich Communication Services (RCS), making it possible for other apps and companies to adopt the technology.

The launch of fake call detection was announced alongside other updates from Android, including a new Google Photos feature that lets users mix and match outfits and try them on virtually. The new “wardrobe” feature catalogs the clothes you’re wearing in your photo library by turning them into snapshots you can browse on your phone. The feature is rolling out next week to eligible users in the U.S., India, and Brazil with Android 10+.

Additionally, Google Play Books is getting a new “Catch me up” feature that lets users jump back into a story with a recap. Users can also highlight a passage to ask questions. These features are rolling out today for select English titles.

Google is also making it possible to search entire outfits with its “Circle to Search” feature. Now the feature will be able to find every item in an outfit at once, getting rid of the need to search piece by piece. This update is now available on all Android 14+ devices that have Circle to Search.

As enterprises troubleshoot their AI systems, governance has emerged as a key challenge. Some are taking a dual approach: One model to handle incoming queries, while a second model keeps the first one from getting into trouble.

That’s the premise of ZeroDrift, a new AI compliance service that announced a $10 million seed round on Tuesday. (Investors include a16z Speedrun, Reign Ventures, PitchDrive Ventures, and U&I Ventures, among others.) The company deals entirely with the second part of the system, sitting between AI models and end users to flag and replace any messages that might present a compliance problem.

It might seem strange to build an AI system to correct other AI systems’ mistakes — but ZeroDrift’s correction system has a few architectural advantages over the models it will be correcting. The system is triggered by conventional programs that deterministically apply known compliance standards like SOC 2 or GDPR. The LLM only comes in once a message has been flagged, rewriting a compliant version of the same message.

“We’re able to identify deterministically, what are all the regulated areas, what’s the violation that’s being broken, and then we have LLMs that can do the rewrites,” Aroomoogan says.

Critically, the whole system can be run with less latency and more reliability than a conventional LLM. It’s the company’s primary advantage over big labs like OpenAI and Anthropic, which are often already present in the underlying system.

The most obvious use case is for AI chatbots, which are already deployed in front of consumers where there can be serious consequences for rogue answers. But Aroomoogan sees the total addressable market as much larger, potentially spanning AI-generated messages that human beings never see, generated only within automated systems. So far, it’s a relatively small market — but it’s one that will grow as AI proliferates.

If the fundraise is any indication, there’s a lot of pent-up demand for that product. “It was probably the fastest fundraising I’ve done in my life,” CEO Kumesh Aroomoogan says, crediting Andressen Horowitz’s help structuring the seed round. “We closed within three weeks, and we will be oversubscribed by 3x on the amount.”