The FBI has issued an urgent cybersecurity warning about a powerful artificial intelligence-driven hacking framework that allows cybercriminals and state-aligned hackers to infiltrate corporate and government networks at unprecedented speeds, potentially threatening businesses and critical infrastructure across the United States. MORE: bit.ly/4zMSsVX
Visa and JPMorgan Chase have landed on the hit list of a ransomware gang, which is giving the financial giants just 48 hours to negotiate before dumping their stolen data online. cnews.link/visa-jpmorgan-dat…
I'm writing day 30 from Burbank Airport, because every United flight was grounded today. I'm hoping a Southwest flight gets me home to SJC.
Day 30 of 30 with @boardyai. For a month I've been posting about agents: how you know which one you're dealing with, whether to trust it right now, which actions should cost a person's attention, and what a board can actually oversee. On a normal day that can sound abstract. It sounds a lot less abstract from a gate full of people rebooking.
Everyone here is depending on systems they've never seen, on the assumption that someone else checked them. Agents are about to inherit that same assumption. More of the infrastructure people rely on will include agents making calls that nobody reviews one at a time.
Here's what I'd keep from the 30 days:
Identity comes first. On day 23 the Agent Name Service became a Linux Foundation project, so an agent can carry a verifiable name.
A track record has to follow that name, so trust doesn't reset at every handoff. That's why we bind TrustScore and AgentCert to the ANS identity.
Runtime checks should scale with reversibility. An action you can roll back can run on its own. One you can't should go to a person, along with the agent's reason for asking.
Boards need evidence, not reassurances written by the teams they're meant to be checking.
Airlines can't treat resilience as optional, and neither can anyone deploying agents. Today was a good reminder of that.
At Burbank Airport, hoping to take off on Southwest to SJC ✈️ A tough day for United, Microsoft, and everyone affected. Today shows how much modern travel depends on critical tech infrastructure. Resilience and assurance aren’t optional. @trustmodelai
Your CISO isn't the bottleneck. The real AI risk? Not asking the right questions.
10 questions every board director should ask. Question #2 changes everything.
Join us Dec 15: trustmodel.ai/bod-risk#AIGovernance#AIRisk@rameshchitor
Robert Pattinson is the new James Bond! 🍸
From Twilight to The Batman to 007, few actors take bolder risks. Add Denis Villeneuve directing, Amy Pascal and David Heyman producing, and Steven Knight writing, and this could be the boldest Bond yet.
Shaken, not stirred. 🎬 #JamesBond #007
An AI hiring audit that only checks who got the offer is inspecting the last door in a building most applicants never got past.
Day 29 of 30 with @boardyai. If a bias audit is going to hold up when someone challenges it, it needs a few things most reports skip.
Group-level impact ratios at every stage of the funnel: screening, ranking, interview, offer. A clean ratio at the offer stage can sit on top of a skewed one at the first screen.
A significance test on each ratio. In a small applicant pool a gap can be noise, and the first thing a reviewer will ask is which one it is.
Calculations someone else can rerun on the same data and get the same answer. If only the vendor can reproduce the number, nobody else can check it.
And a date on the whole thing. Change the model, or let the applicant pool shift, and last quarter's audit describes a system that no longer exists. Reassessment should run every time either one changes.
Most agent controls are set once at deploy time: this tool yes, that tool no. Then the agent spends months making calls that permission list never imagined.
Day 28 of 30 with @boardyai. A permission is a gate. It says nothing about whether this particular action, in this context, should go ahead right now. That matters once you move from models that respond to agents that act. A bad answer can be ignored. A changed record or a sent message is already out in the world.
The runtime policy layer we built at TrustModel checks four things before an agent acts. What's the context? What's at stake? Can it be reversed? And if the agent escalates to a human, what's its reason?
Reversibility does most of the work. An action you can roll back can run on its own. One you can't should cost a person's attention, and the agent should be able to say why it's asking. An escalation with no rationale just hands the problem to someone else.
Agents are cheap and humans are expensive. A good check spends the expensive part only where an action can't be taken back.
Karl Mehta asked a room of 250+ CEOs and directors in Singapore how many of them use AI agents at work. By his account, the real answer was all of them.
Day 26 of 30 with @boardyai. Put that next to the figures from day 21: 83% of S&P 500 companies disclose AI as a material risk, and only 2.7% of their directors have AI expertise.
That means directors are using agents and governing them at the same time, mostly without training for the second job. Using an agent every day tells you it's useful. It doesn't tell you whether it behaved consistently last quarter, or what it did when the data in front of it conflicted.
A director who uses agents can still only sign off on what the board pack shows them. Most packs carry assurances from the teams being overseen, not evidence.
That evidence can be gathered. An agent shows measurable inconsistency before it makes a bad decision, and that measurement is what TrustModel produces and what belongs in front of the board.
On Dec 15, Ashley Carr of @DLA_Piper and I are working through what "reasonable oversight" of AI actually requires.
Perplexity (@perplexity_ai) CEO Aravind Srinivas (@AravSrinivas) spent 15 minutes at SGLang Summit (@sgl_project) talking about “Building the Next Computer.”
Some notes from the talk:
00:44 - How it started. Perplexity built a Slack agent in January where employees talked to a bot more than each other. Marc Andreessen texted Aravind about a company where employees mostly talked to a bot instead of each other. “That was basically us.”
01:35 - Why Computer? Almost SuperAgent or MasterAgent, but he chose Computer, a word that once referred to people doing math for astronomers.
04:20 - Search as code. Fetch, chunk, embed, retrieve, all exposed through a Python SDK so agents can build their own search pipelines.
08:15 - SPACE. Perplexity now runs millions of concurrent sandboxes every day. The platform is open source, including fixes from red-teaming.
12:22 - Local + cloud. A fine-tuned Qwen 27B scored 59.6% locally on 89 tasks; calling Opus 5 only when needed raised it to 73%.
13:20 - Lily. Built by an intern who joined full-time, this Apple Silicon inference engine beats MLX-LM by ~22% on prefill and ~30% on decode on an M5 Max.
He ended by telling everyone they're still hiring.