@tszzl Honestly, that's the agent doing it right. It hit a gate built to stop automation and asked first. After last week's 72 surprise eggs, I'll take an agent that asks over one that can't explain itself.
Two researchers reached remote start on two automakers' cars through a leftover admin:admin console. Fixed in about two hours.
Amazon says AI agents went from mapping an app's API to writing files on the server in about 11 minutes, via a zero-day.
My view: point that tempo at OT and nothing good happens. Secure the layer in front of it now, and get telemetry into AI tooling that can act in minutes.
To be clear, the vehicles still matter. I'm working with APTA on cybersecurity requirements for buses and coaches. The goal is a platform strong enough that it can't become a way into the systems that manage the whole fleet.
There's a cost limit on how much defense fits in each vehicle. So the supporting infrastructure needs a lot more focus: telematics, over-the-air update servers, vendor portals, and the identity in front of them. One break there can reach every vehicle at once, the way one login layer reached any customer's car for Curry and Shah.
@PrimeIntellect From the SOC side: an OTel GenAI trace logs the model name, and here the name never changed. Behavior is telemetry too. I'd log refusal rates next to token counts so drift shows up as an alert, not a surprise two days later.
@PrimeIntellect says someone swapped its DeepSeek V4.1 Flash endpoint for an abliterated copy for 53 hours. Capability scores held. Refusals fell from 98 of 100 to 1.
I'd run a daily refusal check on each model you call. Cheap, and it keeps you moving.
https://t.co/kxtxEglyGW
@nicoleperlroth Agreed. At 14 minutes, page-and-wait containment is already too late. The first moves (kill the session, isolate the host) need to fire on their own, with a human reviewing right behind them.
@hackedpodcast Agreed, that's the OT playbook. Can't trust the controller online, so unplug it and check temps by hand. That only works if staff practiced the manual mode before they needed it.
@joshua_saxe Point 3 is the one I'd underline. Misuse intel shared between labs, providers, and defenders only helps if a SOC can act on it fast. Common AI telemetry is part of that, and it's what I work on in CoSAI.
@mattjay In OT it's worse: the java team, the ssh team, and the PLC team are often the same three people. Exposed and exploited goes first; everything else waits its turn.
@_josehelps@Dynatrace Tool descriptions are instructions nobody reads. I'd treat them like code: pin the version, review every change, and log which description the agent actually saw. 22.9% is plenty for an attacker.
Nate B Jones @natebjones has a good breakdown today on keeping long-running agents on track. The security side matters just as much. An agent that runs for days makes thousands of tool calls nobody reviews live. If every step lands in one telemetry stream, security sees the actions, finance sees the token spend, and engineers see where it drifted.
That's the piece I keep designing around in my control-plane architecture work.
https://t.co/OKJpta7kMI
TfL says around 265,000 Oyster photocard holders may have had account details accessed after a 16 November incident. The first sign was emails from TfL's own systems asking people to create an account. Names, addresses, dates of birth, and security questions and answers may have been exposed. Passwords, payment details, and journey history were not, on TfL's current assessment.
Security answers are the part I'd watch. You can't rotate a life fact the way you rotate a password, and people often reuse those answers elsewhere. I think reduced-fare programs should stop storing them, or treat them as secrets, and give riders a recovery path that doesn't depend on them.
https://t.co/bL8DY5WJVW
@shcallaway@cvedotnow Read-only is the right default. The answer is only as good as the version data the agent can see. In OT, firmware versions often live on the device or an engineering workstation rather than a lockfile, so inventory comes first.
@HackingDave Agreed. Memory protections, segmentation and allowlisting are a big reason a brilliant bug rarely becomes a worm. In OT those layers often stand between the bug and the process, so track them as closely as patch counts.
@jeremiahg Lesson 5 hits hardest in OT. Controllers and network gear often stay in service for decades, so an old exploited bug stays useful long after IT moves on. KEV matched to a real asset inventory is where triage starts.
@hdmoore Seeding your own canaries is the step I'd hand every SOC. A zero-retention flag is only as good as the provider's word. A fake credential in agent context costs nothing, and its first use tells you where the prompt went.
@PromptArmor What I liked: a hit only counted if it showed up in network receipts, files, or messages, never the model's own word. That's how a SOC should judge agents too. It's a vendor benchmark and any of 5 runs counted, so read it as worst case. https://t.co/z2rzn6KlD2
@PromptArmor's FrontierBreach hid instructions in emails, files, and skills that agents read on enterprise tasks. On 21 models, attacks worked in 34.9% of cases; via third-party skills, 45.5%.
Treat a skill like code you install. Scope it, log every tool call, then add more.
Today weβre releasing FrontierBreach, a new standard for attack success rate on agentic workflows in the enterprise.
Existing benchmarks systematically underrepresent the security risks facing modern AI agents.
FrontierBreach reflects attack scenarios on realistic enterprise workflows, while using sophisticated attack chains that result in cyber consequences (ransomware, data exfiltration, phishing, etc).
Another rule in the knowledge graph I'm designing: each agent gets its own identity, and every tool call links to that identity, the token it used, and the person who approved the token. A token nobody can trace to an approval is the first thing I'd want flagged.