Author: Anthony John Padilla

Companies are racing into model-driven services, software agents, and automated development tools while leaving 99.9% of fixable vulnerabilities untouched, according to fresh industry estimates.

More than half of organizations already use software agents in production, and about the same share use them to build custom apps. The rollout is moving faster than patching, access controls, or secure defaults. Model-driven platforms surface roughly 25 confirmed flaws a day. Teams fix about 1.5 of them, and deployment still takes 3 to 5 months.

That leaves companies open to the basics going wrong: missing encryption, exposed interfaces, and permissions that reach too far. The risk gets worse when cloud-hosted models and software agents tie into internal data, third-party tools, and production workflows. Attackers are using the same automation for phishing, adaptive malware, and real-time exploitation. One forecast puts tech-driven cyber incidents in 2025 at 28 million, up 72% from the year before.

If you run model-driven services, software agents, or automated development tools, pay attention to this. Anthropic, Amazon Web Services, IBM, and Microsoft are backing the Akrites coalition, and the U.S. Treasury Department and the Cybersecurity and Infrastructure Security Agency (CISA) have both issued guidance. But gaps across Africa, Latin America, and Southeast Asia, hiring shortages, and too much reliance on machine output mean the security groundwork is still falling behind.

You can find guidance online from CISA, the U.S. Treasury Department, and the Akrites members.

Anthropic launched its drug discovery program today. It’s a clear move into neglected-disease research, and it also introduces Claude Science, the company’s workbench for genomics, proteomics, and single-cell analysis.

Claude Download

The company is moving past selling tools to labs and into preclinical discovery itself. Anthropic’s argument is that this stage fits AI unusually well: biology data is abundant, simulation and ranking can happen quickly, and failures cost less here than they do later in the process. If that plays out, early medicine development could speed up, and the work could feed back into better life-sciences products.

There’s a different business model here too. Anthropic says its public benefit corporation structure gives it room to put patients’ needs ahead of short-term returns. That comes at a moment when Novartis CEO Vas Narasimhan has said these tools could cut timelines from about 12 years to 7 or 8, while pushing success rates from roughly 8% to 16%. Anthropic also says Claude Science connects to more than 60 databases. In its early examples, a University of California, San Francisco researcher spotted viral contamination in minutes after it had been missed for a year, and another analysis covered 100 rare genetic diseases in under an hour and surfaced 32 follow-up candidates.

If you follow AI in biotech, this deserves a close look. Discovery is still only the first part. Clinical trials cost come after that.

You can get Claude Science through Anthropic’s research offering, and Anthropic’s new drug program is starting with preclinical work.

At Adobe Summit, Adobe laid out a bigger plan for its product lineup: one connected system that links Photoshop, Adobe Express, Adobe Acrobat, and Adobe Experience Cloud, covering content creation, workflow management, and customer experience delivery.

On the creative side, Photoshop and Adobe Express are picking up smarter editing tools. Acrobat is heading in a different direction, toward documents that feel more interactive and are easier to search, summarize, and act on.

What stands out most is Adobe’s effort to connect those apps with business software for customer data, automated campaigns, and personalized websites and apps. That’s a strategic bet at a moment when Adobe is dealing with heavier competition from Canva on the design side and Salesforce in enterprise software.

Disney gave the clearest example Adobe had. The company walked through a scenario where someone spots a piece of media or a promotional offer online, starts planning a Disney World trip, uses the app while they’re there, and then gets follow-up recommendations afterward. Adobe is now selling that same full-path approach to other industries too, including theme parks, streaming, retail, and travel.

If your work depends on Adobe’s creative tools or its business software, this broader push deserves attention. Adobe is trying to influence what happens after an image is created or a document is sent, all the way from an online ad to a hotel booking to a family’s next day at Disney World. You can follow the latest announcements at Adobe Summit.

Microsoft has laid out a new attack technique called AutoJack. In its write-up, the company shows how a malicious webpage could use a web-enabled agent as a go-between, reach a local Model Context Protocol (MCP) service, and kick off arbitrary processes on your computer without asking for any extra clicks.

What this comes down to is a confused-deputy problem. Once a web-enabled agent can browse the web and also talk to privileged local services, localhost isn’t a boundary you can safely assume will protect you.

Microsoft says the proof of concept relied on three separate issues lining up at once: inherited localhost identity that slipped past an origin allowlist, MCP WebSocket endpoints with no authentication, and URL-based server_params that were decoded straight into process-launching logic, with no executable allowlist in place. Put together, that gave the webpage a way to tell the local service to run code. Microsoft also says this wasn’t a browser bug. It was the agent’s trusted position being used against it.

If you’re building or deploying web-enabled agents, this should get your attention, because the same pattern could easily turn up in other frameworks that connect web access with local tools or developer services.

Microsoft says the vulnerable AutoGen Studio MCP code only appeared in development builds and was fixed before anything reached a public PyPI release. Even so, it still points to the broader risks around prompt injection, agent hijacking, data exposure, and the need for zero-trust controls, sandboxing, and runtime monitoring. If you want the full breakdown, Microsoft’s research goes into the details.

Nvidia has unveiled ENPIRE, a research system built with Carnegie Mellon University and the University of California, Berkeley. The idea is simple enough: train robots on real hardware instead of only in simulation, and do it with very little human involvement.

Nvidia says ENPIRE runs a fully autonomous loop from start to finish. It resets the workspace, tries a strategy, checks the result, and updates its policy.

In the company’s testing, eight robots reached success rates as high as 99% on manipulation tasks that would normally need manual resets, hand-built reward functions, and close engineering oversight.

Those tests covered Push-T, sorting pins into a box, cutting or fastening cable ties, pin insertion, and assembly-style work like installing a GPU into a motherboard.

Nvidia also says ENPIRE can generate reward functions and success checks from only a few minutes of success-and-failure video. It does that by using visual alignment, gripper height, estimated force, and views from two cameras, while keeping response times under 150 ms.

If you follow robotics, this one deserves a look.

After a lightly supervised initial setup, Nvidia says eight dual-arm YAM stations can share code through Git, read papers, test hypotheses, and decide whether behavior cloning or reinforcement learning makes more sense.

In one example, Nvidia says ENPIRE reached 100% pin insertion faster than a human-in-the-loop baseline. The company says ENPIRE builds on Nvidia’s Eureka, which beat human-written rewards on more than 80% of tasks, with gains above 50%, and Nvidia plans to open-source ENPIRE alongside Project GR00T and Nvidia Isaac.