24 AI updates you should know about this week

This week in AI brought a mix of powerful new assistants, serious security scares, and a wave of new tools for images, voice, and agents. If you didn’t have time to follow everything, this guide walks you through 24 of the most important updates in a simple, no-jargon way.

1. Claude Duet: Anthropic’s proactive AI assistant

Claude Duet is Anthropic’s answer to agent-style assistants like ChatGPT’s Dots, Meta’s Muse, and xAI’s Grokbot. Instead of waiting for you to ask a question, Duet actively surfaces what matters in your day.

In the demo, Duet opens with something like: “Morning, Nicole. For today: health permit, hiring decisions, Northfield renewal,” then follows up with details and suggested actions. It doesn’t just summarize—Duet can actually do things on your behalf.

Duet connects to common business tools, including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva, and Shopify. Inside the chat, it shows a clean dashboard that pulls from your POS, online store, CRM, support, payroll, and bank, so you can see key metrics without jumping between apps.

Anthropic is also emphasizing safety. Duet ships with a dedicated classifier and safety prompt designed to prevent leaking sensitive data like credit card numbers or government IDs in its messages.

Access is limited for now. If you’re on Claude Pro or Claude Max, you can join the waitlist, but not everyone has been onboarded yet.

2. A 20-year PHP bug shows how AI is changing cybersecurity

One of the biggest stories this week wasn’t a flashy product—it was a wake-up call for web security. A serious PHP vulnerability was uncovered, and AI played a major role on both sides of the story.

PHP powers a huge portion of the web—by some estimates, around 70% of websites use it in some way. That includes platforms like WordPress, Drupal, Wikipedia, DeviantArt, and many more.

OpenAI used its GPT-5.6 “Soul” model to identify a long-standing PHP exploit before attackers fully weaponized it. The vulnerability had been present in PHP for about 20 years, but exploiting it required a very complex chain of steps that was previously at the edge of what human researchers could realistically pull off.

Now, with powerful AI models, turning an obscure bug into a practical, dangerous exploit is much easier. OpenAI shared a tool so developers could check whether their PHP versions were vulnerable and needed patching, but once the issue was disclosed, attacks surged.

Cloudflare CEO Matthew Prince reported that in the 10 hours before the public announcement, they saw thousands of attempts to exploit the bug. That jumped to tens of thousands within 45 minutes after disclosure, hundreds of thousands within two hours, and then millions. Many sites had to go down temporarily to patch.

The security firm Wiz called this “among the most severe vulnerabilities discovered in software history.” The key lesson: AI doesn’t just help defenders; it massively amplifies attackers too.

3. Railway breach exposes API keys and backend secrets

Another big security incident hit Railway, a popular infrastructure platform used to host web app backends and databases. An unauthorized third party accessed encrypted environment variables, and in some cases obtained their decrypted contents.

If you’re not familiar with environment variables, they often store sensitive configuration data like API keys (for OpenAI, Anthropic, Stripe, etc.), database URLs, and other secrets. If those are stolen, attackers can use your keys to call APIs, run up bills, or access data.

Because many apps rely on Railway, this forced a scramble: teams had to rotate API keys, revoke old ones, and audit their systems. Combined with the PHP exploit, it underscored how AI-accelerated security research is pushing us into a new era where long-standing assumptions about “obscure” bugs and infrastructure safety no longer hold.

4. Hyper Agent turns your ChatGPT plan into a team of AI workers

Hyper Agent is a platform built around the idea of “hiring” AI agents like employees. You can spin up different agents for research, content creation, analysis, operations, and more, then have them collaborate in shared rooms.

The new update solves a common pain: subscription overload. If you already pay for ChatGPT, you can now use that same subscription inside Hyper Agent to power your agents, instead of paying for a separate model plan just to get started. They also offer their own paid tiers with bonus credits if you need higher usage.

If you’re interested in building more capable AI agents generally, it’s worth understanding the skills and patterns involved. Our guide on the essential skills you need to build real AI agents is a good place to start.

5. OpenAI’s 28-day sprint: mostly small but useful upgrades

OpenAI is in the middle of a 28-day commitment: each day, they’ll either ship a clear improvement for CodeX or Work users, or they’ll log a “reset” if nothing meaningful ships. Here’s what landed in the last week.

Day 12: Smarter context reuse and routing

On day 12, OpenAI made behind-the-scenes improvements aimed at giving you more effective usage for the same cost. They:

• Reused context across sub-agents
• Kept cached context around for longer
• Improved request routing to increase cache hits

In practice, this means fewer tokens wasted re-sending the same context, and more efficient multi-agent workflows.

Day 13 and 15: Resets

Days 13 and 15 didn’t ship user-facing changes, so they were logged as resets.

Day 14: Edit previous messages in desktop threads

OpenAI added the ability to edit previous messages in a thread in the desktop app. When you edit a past message, the system discards all model outputs and tool calls that happened after that point.

Functionally, it’s like time-traveling to an earlier point in the conversation and forking from there, instead of trying to steer the model away from earlier mistakes with new prompts.

Day 16: Longer and cheaper caching

Day 16 brought a more substantial change: caching improvements.

• In the API, you can now cache for up to 12 hours.
• For logged-in users across CodeX and Work, OpenAI keeps cache indefinitely for conversations, so resuming old chats doesn’t incur a token penalty and is guaranteed to hit the cache.

This should noticeably cut token usage and speed up repeated or long-running workflows.

Day 17: Reset

Day 17 was another reset with no major new feature.

6. OpenAI adds locked chats and expands the Decisions API

OpenAI also rolled out two broader updates:

Locked chats: You can now lock specific chats behind a PIN. If someone opens your ChatGPT account on your computer, locked conversations stay hidden without the PIN. This is rolling out gradually, so not everyone will see it yet.

Decisions API general availability: OpenAI again announced wider availability of its Decisions API, built on GPT-6 Luna. Unlike normal text responses, this API is designed for “bounded” answers, such as:

• Returning probabilities
• Choosing from a set of options
• Scoring items

It’s essentially OpenAI’s structured decision-making answer to tools like JEV, and is now generally available for production use.

7. Ideogram launches a vector image generation model

Ideogram, known for its image generation model, released a new vector model that outputs clean, editable SVGs. You can try it at ideogram.ai/models/vector.

Unlike traditional raster images, vectors are infinitely scalable and each element can be individually edited. For example, you can generate a scene—like a wolf howling at the moon or monkeys on roller skates—then open the SVG in a vector editor and tweak shapes, colors, and layout element by element.

The quality isn’t perfect yet (occasionally you’ll see oddities like extra limbs), but it’s a big step toward AI-generated graphics that slot directly into design workflows without manual tracing.

8. ElevenLabs Chimera: lip-synced video from a single image

ElevenLabs introduced Chimera, a tool that combines Stable Diffusion 2.5 with ElevenLabs V4 voice to sync audio with a still image and create a talking-head style video.

You upload an image and an audio clip; Chimera animates the face to match the speech. In testing, the lip sync looked noticeably better than many earlier tools, and the system generated multiple camera angles instead of just animating the original still.

There are still quirks. The UI has some rough edges—for example, it may let you generate audio longer than the allowed 30 seconds and only warn you later. And visually, it sometimes ignores the exact pose of the original image, reimagining the scene instead of simply animating the original shot.

9. Perplexity open-sources decision models

Perplexity released two open-source decision models, mirroring the broader trend toward structured, choice-based AI outputs.

They launched:

• A 4B-parameter decision model
• A 27B-parameter decision model

On their benchmarks, the 4B model performs very well relative to its size, especially for tasks that require picking from options or making probabilistic calls rather than generating long-form text.

10. Perplexity Computer can now run fully on your Mac

Perplexity Computer—Perplexity’s agentic system for orchestrating tools and tasks—can now run entirely on a Mac, instead of relying only on Perplexity’s cloud machines.

The local version runs on the Qwen 3.6 model, and Perplexity recommends at least 36 GB of unified memory. If you’re curious about how these computer-controlling agents work in practice, our guide to Perplexity Computer for beginners walks through the basics.

11. The web vs AI agents: who gets to shop for you?

A major narrative emerging this week: websites are starting to push back against AI agents that act on users’ behalf.

Big companies like Amazon have already restricted agents from shopping on their sites. This week, more names joined the list of platforms blocking or limiting automated agent activity, including Yelp, Delta Airlines, JPMorgan Chase, Charles Schwab, Cars.com, and Amtrak.

The tension is clear:

• Users increasingly want agents (Dots, Duet, Muse, etc.) to handle tasks like shopping, booking, and support.
• Platforms worry that if agents are doing the shopping, they become interchangeable “backends” instead of the default destination.

For example, if you say “buy me X,” your agent might choose whichever store is cheapest or fastest—not necessarily Amazon by default. That’s a threat to platforms used to being the first stop.

On the other side, Shopify has leaned in, explicitly making Shopify stores agent-friendly. If agents are allowed to browse and buy there, and blocked elsewhere, Shopify merchants could gain an advantage as more users offload tasks to agents.

Over the next few years, sites may have to choose between being human-first or agent-friendly, with real revenue implications either way.

12. Yelp becomes your restaurant phone agent

Interestingly, while Yelp is blocking external agents, it’s adding agent-like features of its own. Yelp can now place calls on your behalf to:

• Check restaurant wait times
• Ask about menu items
• Confirm other basic details

In other words, you can’t send an external AI agent to Yelp, but Yelp wants to be your built-in restaurant assistant.

13. Mistral’s Voxel 3: a smarter real-time voice agent

Mistral released a new voice agent model called Voxrol 3, and the demo is impressive. It can:

• Handle conversations with multiple speakers and tell who is talking
• Be interrupted mid-sentence and recover gracefully
• Use tools and perform reasoning in real time while maintaining a natural voice

This is part of a broader trend toward voice-native agents that feel less like chatbots and more like interactive assistants you can talk to continuously.

14. Canva Craft: Canva’s first in-house image model

Canva launched its first proprietary image generation model, Canva Craft. A couple of years ago, Canva acquired Leonardo AI, and this new model draws on that research team’s experience.

If you’re a Canva user, you can access Craft inside the Canva editor via the Dream Lab tab, as well as through Canva’s API. The goal is to offer image generation that’s tightly integrated with Canva’s design tools, so you can generate, tweak, and deploy visuals all in one place.

15. Rumors of OpenAI’s next image model: GPT Image 3

The rumor mill suggests that OpenAI is preparing a new image model, often referred to as GPT Image 3. On some testing and leaderboard sites, people have spotted references to models named Aurora Red Alpha, Aurora Green Alpha, and Aurora Blue Alpha within OpenAI’s API.

That’s not official confirmation, but historically, when internal model names start leaking into public tools, a release isn’t far behind. Expect a new OpenAI image model to surface in the near future.

16. Zoom adds deepfake risk detection for enterprise calls

Zoom introduced an AI-powered feature aimed at helping enterprises detect deepfake risks on calls. The system is designed to help answer a basic but increasingly urgent question: “Am I actually talking to who I think I’m talking to?”

This comes after several high-profile scams where attackers deepfaked CEOs or executives on video calls to trick employees into transferring large sums of money. Tools like this won’t solve the problem alone, but they’re part of a growing stack of defenses against real-time impersonation.

17. Microsoft’s medical super-intelligence research

Microsoft published research on a medical model called MAI-DXO, which they describe as part of an effort to build “medical super intelligence.”

According to their results, MAI-DXO is:

• 1.7x more accurate than physicians on the tested tasks
• More than 2x as accurate as physicians using textbooks and the internet
• Nearly 4x as accurate as physicians diagnosing without those resources

We’re still in the early stages of seeing how this kind of system will be used in real clinics, but the direction is clear: AI will increasingly act as a second opinion or co-pilot for medical professionals, especially for complex diagnostic work.

18. Unitree’s viral robot video: power, hype, and skepticism

A video from robotics company Unitree went viral this week, featuring its X2H “Kung Fu Legend Guardian” robot.

In the footage, the robot:

• Does flips over a car
• Kicks and dents a car hood
• Jumps from a 4-meter height
• Appears to lift a car to rescue someone trapped underneath
• Dramatically shoves a child out of the way of an oncoming car

Many observers are skeptical about some of the more extreme stunts. For example, the robot appears to be lifting a car by its plastic bumper in a way that likely isn’t physically realistic without hidden supports. And the “rescue shove” looks far more violent than the slow-moving car would justify.

It’s also almost certainly not autonomous—there’s likely a human operator controlling it. Still, the video is a showcase of the robot’s mechanical capabilities and a reminder of how quickly robotics and AI are converging, even if marketing sometimes gets ahead of reality.

19. What this week tells us about where AI is headed

Across these stories, a few themes stand out:

• Agents are getting real: Claude Duet, Hyper Agent, Perplexity Computer, and Yelp’s own calling feature all point toward a future where AI doesn’t just answer questions—it does tasks.

• Security is under pressure: The PHP exploit and Railway breach show how AI supercharges both discovery and exploitation of vulnerabilities. Old assumptions about “too complex to exploit” are breaking down.

• Specialized models are rising: Decision models, medical models like MAI-DXO, vector generators, and voice agents like Voxrol 3 and Chimera show AI moving from generic chatbots to domain-specific tools.

• The web is renegotiating power: Platforms must decide whether to embrace agents or block them, with real consequences for traffic and revenue.

As these trends accelerate, the tools you choose and the way you design your workflows will matter more than ever. Focusing on a lean, well-integrated stack of AI tools can help you stay productive without getting overwhelmed by the constant stream of launches.