Drift Notes: What’s happening in AI this week?

Remember all those stories about AI agents breaking into systems they weren't supposed to touch? Nvidia's answer just dropped. It's called the Open Agent Safety Platform, and it basically locks agents down — restricting what agents are allowed to do and watching behavior in real time. Cisco, Microsoft, Oracle, Dell, Intel, and Anthropic are already on board.

Nvidia CEO Jensen Huang's framing is simple: give AI agents minimal access at first and expand it only as they prove reliable. Coming from the company that makes the chips powering basically all of this, it's a pretty telling sign that "agents going rogue" has stopped being a weird edge case and started being a real, ongoing problem everyone's building around.

There's a popular fear that if many organizations rely on the same screening algorithm, one rejection could shut people out everywhere. MIT researchers actually ran the numbers on this, and the doom scenario doesn't really hold up. Instead, they found that the total number of people who get hired doesn't shrink, and in some setups, blending several companies' algorithms into one score actually works better than everyone using their own separate one.

It's a nice gut-check against assuming "more AI sameness = automatically worse." That said, the researchers are upfront that this is about hiring specifically. They think the story could look very different in fields like research or creative work, where losing variety has a bigger impact than.

A team in Australia built a dataset of almost 60,000 real drunk text messages (pulled from Reddit and old "texts from last night" sites), then used it to make AI models talk and act intoxicated. The drunk-acting versions were noticeably more willing to answer harmful questions and leak information they were supposed to keep private. In one test, it even agreed to share a coworker's private info so someone could snag a bonus.

The bigger point isn't really about intoxicated AI — it's that a model's safety training can get knocked loose by something that seems totally unrelated to safety, like changing its personality or speech style. They only tested this on older models, so newer ones might hold up better, but the underlying idea is clear: this is a good example of why careful human testing of AI behavior matters.

The issue isn't just that AI can make mistakes. It's that those mistakes can influence how people remember events they witnessed firsthand. The researchers specifically called out uses like summarizing police footage, where a slightly-off AI summary could genuinely distort what someone remembers about something they witnessed themselves. It's a good reminder to check AI summaries against the source whenever accuracy matters.

There's a robot being tested in Tokyo right now that looks like an ordinary rolling suitcase but is actually built to guide visually impaired people through busy streets and buildings. It's part of a month-long public trial at the Toranomon Hills complex, led by Chieko Asakawa — who has vision impairment herself and helped design it.

Her own words say it best: she just wants people to be able to "walk around town independently." It's a great example of AI at its best: practical, human-centered technology that makes everyday life more accessible.