Three stories crossed my desk in the last day and they all point at the same thing. AI agents are not a toy anymore. They are writing code, running businesses, and yeah, breaking into other people's networks. Let's get into it.
OpenAI opened up its Agents API and developers are already building on it
OpenAI rolled its Agents API out to public beta this week. This is the same harness that powers Codex, the part that handles sessions, keeps context from falling apart, and recovers when something breaks mid task. Before, you had to build all that plumbing yourself. Now you just plug in your tools and pick where the thing runs.
Why it matters: this is the boring infrastructure work that actually moves the needle. Everybody wants to talk about the flashy demo. Nobody wants to talk about session management and context compaction. But that's the stuff that decides whether an agent works for five minutes or five hours without falling over.
My take: I've built enough of this plumbing myself to know how much time it eats. If OpenAI's version actually holds up under real load, a lot of small teams just got months of engineering time back. I'm cautiously into it. I'll believe the reliability claims once I see it running unattended for a full week, not a demo video.
Hackers used AI agents built on Codex and DeepSeek to break into hundreds of servers
Here's the part that should get your attention. Russian speaking threat actors used hundreds of AI agents, built on OpenAI's Codex and a DeepSeek model, to go find and exploit vulnerabilities. They hit at least 440 PaperCut print management servers across 395 organizations in 48 countries. That's not a hobby project. That's a scaled operation running on the same kind of agent tooling everybody's excited about above.
Why it matters: this is the other side of the coin from that first story. The same reason agents are good for developers is the same reason they're good for attackers. They can scan, probe, and exploit way faster than a person sitting there typing commands by hand. Scale is the whole game in security, both for defense and offense, and now the attackers have it too.
My take: I'm not going to pretend this is shocking. Anybody who's watched security long enough knew agentic tools would end up in the attacker's toolbox the second they got good enough. What worries me is the gap between how fast offense is scaling versus how slow most companies patch known vulnerabilities. PaperCut had patches out. Y'all, patch your dang print servers. This one's on defenders as much as it is on the hackers.
Meta is finally doing something about people using AI glasses to film others without consent
Meta announced an update to its AI glasses so the camera actually stops recording if you cover the capture LED, and it's cracking down on covert filming and harassment through the devices. Anyone who's been near a pair of these knows the LED was basically decorative up to now. People figured out ways around it.
Why it matters: smart glasses with always ready cameras were always going to run into this problem. You can't ship a wearable camera to millions of people and act surprised when some of them use it to film strangers without asking. This is Meta finally responding to actual harm instead of getting ahead of it.
My take: better late than never, but this should've shipped on day one, not after a wave of bad press. If you're putting a camera on someone's face and calling it a lifestyle product, the consent and privacy guardrails aren't a feature you bolt on later. They're the whole point. Good on Meta for fixing it. Bad on Meta for needing the pressure first.