unauthenticated remote code execution

I spent the last month obsessed with finding colors that can't be displayed on a conventional screen. This is what I found. moultano.wordpress.com/2026/06/19/w...

Fixing all jailbreaks deterministically is simply not possible with what we currently know about LLMs. It's a "no, no, no, do what I *mean*" problem. In most cases, we can't even define what a jailbreak is, it's a case of I know it when I see it.

Fifteen years after leaving my postdoc I'm still getting email spam for circadian monitors and custom DNA synthesis.

This is the World Cup content I'm here for.

Since recent events seem to have dragged AI powered biosafety back into the chat, thought I'd take the excuse to repost this. IYKYK.

Best advice I ever got about 'ideating' was to sit down and grind out a ridiculous number, a hundred or more, and then throw away the first half of the list. Those are the ones everyone has, but they're also the scaffolding to get you to the good stuff. AI only gives you the first half.

"To the slop architect, execution, skills, and experience are secondary, bowing at the pedestal of ideas. The fact is, most ideas are ill-thought-out, half-baked, or just plain fucking stupid." perilous.tech/the-architec...

Pay for the things you want to see more of in the world.

2005 called and wants its hostile vendor culture back. Let's take a moment to talk about how “responsible” disclosure would mean vendors don’t give bug hunters the runaround and slow walk fixes, then cry victim when the researcher discloses the bug publicly. 1/

Trying to draw more. Will all be in this thread, mute it if you're here for Very Serious Security Content.

I cannot begin to describe to you how bad I am about promoting my own work, partly because it consumes my entire brain while I'm putting it together, but once it is out in the world, I forget it exists because I have already moved on to the next thing.

The capabilities are the risk. I don't know how many times I have to say this. If your agent gets *any* untrusted input *ever*, then you have to build your agent under the assumption that its output and goals are potentially under the control of an attacker.

Doing some replays, and both of the Hades games are so unreasonably good and replayable. You can see how Hades II polished a lot of bits from the first, but the first still emphatically holds up (and I think still has the best two songs).

Choose your warrior.

I'm just a SecOps and Threats guy, not selling a damn thing, so whenever you talk about AI "democratizing" capabilities to an employee, I want you to keep in mind for me that short of herculean new spends and deliberate frameworks, it's without the dev, QA, monitoring, and responsibility offload.

Where does the “it’s not x, it’s y” come from? It’s not only training data: LLMs are trained to use language to arrive at “answers.” Whatever we’re incentivizing and disincentivizing in our use of language when we resist “sounding like AI” has big implications for students, and for human thought.

gimme the teeth boys and free my soul

i feel like this is a great PSA about the perils of the docker group

I can't express how much I hate having to review LLM-generated slop from 'writers' with no expertise in the area that they didn't bother reviewing at all. Overstated claims, incoherent framing, nonsequiters stuffed into lists, glaring technical errors that even a brief review would have caught.

The statement 'we do/do not understand how LLMs work' almost invariably confuses two very different things. On the one hand, we absolutely can map and describe, in great detail, every single mathematical operation that they perform to generate an is answer. But...

Chilling Effects

An unapologetically angry post about what’s being done to US science:

The onboarding where it *aggressively* pushes you to to connect every single source of health data they could think of, down to searching for doctors records and medications, to Google Health was pretty offputting too.

How to guarantee a speaker gig: Hack the system. Literally

Fascinating story of how IRS enforcement of wealthy tax cheats has been continuously roadblocked since the 70s. "President Richard Nixon promptly installed Donald Alexander—a tax attorney who represented ultrawealthy clients and was openly hostile to enforcement—as his new IRS commissioner."

Seeing another previously-good newsletter start churning out Claudeslop. Disappointed.

"If you want, next I can--" <guttural scream>

For $7 an hour, virtual assistants in the Philippines are using AI to write LinkedIn posts and comments on behalf of Western executives, fueling a "thought leadership" content mill https://restofworld.org/2026/virtual-assistant-linkedin-engagement/?utm_campaign=row-social

I have long wondered what happens if/when someone shouts "Westphalia, motherfuckers!" and takes direct control of the various peering points etc. in their sphere of influence.

I say without hyperbole that this is the most brazenly corrupt action in US Presidential history. That it does not immediately lead to impeachment is a dangerous sign of how far the rule of law has declined.

not now strange under-ice structures, can’t you see there’s enough weird shit going on

It is 9:58am on a Monday and I am already sick unto death of getting Claudeslop full of bullet lists with **BOLD TEXT** and *emphasized placeholders* and busted mermaid diagrams that's all 2-3x as long as it needs to be when I ask for technical documentation and architecture diagrams.

Quoting because I hope someone has tried or will try this.

All that, plus: if you see it as a ridiculous burden to verify a citation, then a) you might not be sufficiently familiar with the literature of the field you're publishing in, and b) you have a fundamental misunderstanding of what citations are even for.

#StreetArt by Peeta Manuel Di Rita in #Delft #Netherland

I suspect we'll ultimately settle at C2, or possibly C3 for values of C3 that exclude "machine god". I think we're at C2 now but I can see how folks could argue for A2. Column 1 is simply demonstrably incorrect, and row B (IMO) remains a very silly belief.

About to become the world's dumbest cash prize ctf.

post some good^H^H^H^H pencilslop

'The Boys' is a *deleriously* stupid show at this point (you can almost hear the writer's room begging for someone to please stop them) but Frenchie saying "You have never danced a day in your life" was a 10/10 line.

AI agents = access groups where once you know how to reach them, the agent can be manipulated to follow instructions no matter where the instructions come from or what they are. This is going to go so badly.

if you are submitting a talk, your goal should be that i can't just copy paste my standard comments at your talk. docs.google.com/document/d/1...

LLM-as-a-judge my dude that is SO 2025 we're doing Agentic Juries no wait we're doing Agent Councils NO NO NO Agent *Congresses* wait: AGENTIC UN. You people are all so cooked. SO cooked. I'm gonna reimplement Venetian elections via agents for maximum incorruptibility.

There's a very old joke in the field (dating back to the 90s at least) When we hire people, it's machine learning. when we do the work, it's logistic regression. And when we raise money, it's AI.

You really don't want AI in your automation pipeline if you can help it. Use an LLM to design a deterministic pipeline, let it go wild and make a really complex one if that's what's needed, but then deploy the thing that's checkable and predictable, not the thing that can change its goal mid-run.

I can't wait for something to change so this can be a boring technology that's used in industry like, idk, spreadsheets or build tools. The constant stream of grifter shit is exhausting.

Help me continue to ignore things like work/life balance:

It's not "signs of AI generated writing" it's IOCs (indicators of Claude).