Choose your warrior.

If you're building internet-facing services on the MCP Python SDK, update to >= 1.10.0 at your earliest convenience. Fixes two unauthenticated network DoS issues (CVSS 8.7) in <1.9.4 and <1.10.0, respectively. nvd.nist.gov/vuln/detail/... nvd.nist.gov/vuln/detail/...

"Give me your tired, your poor / Your huddled masses yearning to breathe free / The wretched refuse of your teeming shore / Send these, the homeless, tempest-tossed to me..." Also: "Truly I tell you, whatever you did for one of the least of these brothers and sisters of mine, you did for me."

If you insist on taking LLM output and rendering it via markdown, at least use image content security policies; it's not perfect for preventing exfiltration but it'll at least make them work for it.

"AI-powered VScode fork" is the new "ChatGPT wrapper" change my view.

Shower thought: the one-two punch of raw empiricism combined with 'state of the art' chasing for AI academic work has diverted a lot of research into pure emprical performance tuning that might have been better spent on theoretical work.

Thought experiment: if you were going to set up a watering hole attack for indirect prompt injections against unknown agents, where would you do it and what sort of injections would you try? I'm not sure about where/how, but a markdown image fetch to a server I control feels like the right payload.

Just constantly asking "what does the bottom turtle look like?" these days.

Meanwhile, on Twitter (not "X"; their words not mine).... (From quick inspection: mostly crypto + telegram scams -- this is about a week's worth)

If you're interested in the security of agentic systems, you're not going to want to miss this talk. @beccalunch.bsky.social will present NVIDIA AI Red Team findings in real world agentic systems, and I'll talk about how the AI Security team helps mitigate them. www.blackhat.com/us-25/briefi...

Here's the thing: You probably already have "Computer Use Agents" running on your corporate machines, with all of the _insane_ enterprise risks they entail. We're just calling them "IDEs" for some reason.

If you're not aware: the MCP inspector tool is a great one to have in your back pocket for poking around MCP servers looking for weaknesses. modelcontextprotocol.io/docs/tools/i...

A friend with a Pixel 9 (Android 15) cannot figure out how to turn off Google’s AI suggestion features, which are even colonizing Signal. They’ve tried every switch they can find. Any ideas?

I think I've said this before, but the one AI risk I now believe that I was genuinely miscalibrated on was how readily people form social attachments to chatbots. Optimizing them for sustained engagement seems likely to be -incredibly- corrosive to the mental health of a nontrivial number of people.

Let’s goooo!

Randomly remembered the saddest thing online

It’s funny because it’s true

"When a metric becomes a target it ceases to be a useful metric." arxiv.org/abs/2504.20879

Tapping the "Models give you what you ask for, not what you want" sign yet again.

New blog post. developer.nvidia.com/blog/structu...

my idiot best friend has cancer and told me to reskeet this and since she has cancer I have to do it she said that to me at a dinner but for real, it's a good read and maybe if you're having weird stuff happening in your ass you should go get checked just in case

I was planning on writing another "how to write a good submission" post at some point once the current review season is over (why is the spring always so crazy?), but at this point I think my number 1 suggestion would be "actually complete your submission. like all of it. actually finish."

I feel like the *real* rule zero of US politics has always been "don't fuck with the money". So now that this has gone far enough that formerly staunch allies are openly musing about dropping USD as a reserve currency, I'm extremely curious to see where this goes.

Fun read on a supply chain attack- www.landh.tech/blog/2025021...

Large language models store vast amounts of knowledge, but how exactly do they learn it? Excited to share my Google DeepMind internship results, which reveal the fascinating dynamics behind factual knowledge acquisition in LLMs!

Pre-coffee hot take of the day: Learning to draw has convinced me that making visual arts by hand teaches you how to *see* and communicate with images, in a way that using GenAI to one-shot something never can. I strongly suspect that the same is true for text/verbal thinking and music/hearing.

"[we] are looking at maybe another 48 hours of intense posting until the fad boomerfies on Facebook while the online vanguard moves on to some other annoying trend" is a sentence that I am extremely upset that I can both parse and know instantly what it's referring to.

OneDrive: "I saved your file automatically for you!" Me: "Ok, but now I can't find it anywhere, where did you save it _to_?" OneDrive: "That's secret."

AI enabled systems are software. You do not need a whole new parallel CVE framework for vulnerabilities in AI-enabled systems. Please stop re-inventing them. I've been in this space for 10+ years and the _only_ new thing (AFAIK) that doesn't map ok onto existing stuff is training data poisoning.

The more I think about it the more I think the 'smart-stupid' framing is Bad, Actually. People are good at things (or not) and we've picked a bucket of academic-ish things, labeled people who are good at them 'Smart', and then -moralized- it.

Chrome just forcibly disabled uBlock origin (as promised like six months ago) and there was a brief moment of "Damn; people live like this?" before re-enabled it. Time to find a new browser I guess.

In a lot of ways, it feels like the LLM "Jailbreaking" subdomain is replaying a lot of the same beats that the "adversarial examples" subdomain ran through about 8 years ago. Similar back-and-forth, similar explosion of techniques with ineffective defenses, similar lack of consistent impact. 1/

I am begging AI Red Teams to stop killing themselves trying to prevent attacks that can be just as easily accomplished by editing client-side HTML. For example:

Severe risk of out-of-work personnel with sensitive knowledge and skills being compromised by foreign intelligence services. But yay look at the doggy clock go! 🤦🏻‍♂️

Two weeks from now the crew of a nuclear submarine are going to see this email and the subsequent one firing them all.

Cartesia.ai distills LLaMa into a state space model with interesting results The distilled version seems to consistently outperform on reasoning tasks, even though it follows a Mamba2 architecture. So, the "reasoning" capabilities seem to emerge from the underlying architecture.

The WARSONG UPDATE for HADES II is here!!⚔️ Face the final confrontation, meet Ares, and more in our second Major Update in Early Access! 🩸Trailer: www.youtube.com/watch?v=yrpG... 📝Patch notes: store.steampowered.com/news/app/114... 🌒Play it today on Steam and the Epic Games Store!

Did I miss a vibe shift, or when did tech folks pivot from being into like Soylent/theanine/piracetam with occasional gray market Modafinil into taking stuff like ayahuasca and ketamine and microdosing LSD? I thought we were interested in keeping our brains between our ears?

Irregular "complaint about OneDrive" post. Love how I have to actively hide my docs from OneDrive's syncing because otherwise one Microsoft product will *actively destroy* the output of another Microsoft product in such a way that it is completely unrecoverable.

Russia-linked hackers have been phishing Ukrainian military with Signal QR codes that look like group invites but instead add a linked device that eavesdrops on messages. Update Signal now to get its fix for a tactic that’s likely to spread beyond Ukraine. www.wired.com/story/russia...

This is maybe small potatoes in the overall scheme of things, but given the likely future state of online consumer protection, just take a beat to consider how much *more* rapacious, exploitative, and predatory online gambling is poised to become over the next four years.

In the coming weeks, a lot of slow-on-the uptake folks are going to begin belatedly realizing Trump is both an inept moron and a dangerous autocrat. Our great civic challenge in these trying times will be calmly welcoming them aboard instead of primal-screaming in their faces.

"Gulf of America" is just a loyalty test. It's a way of checking who is going to abrogate reality for no reason other than a cowardly refusal to stand up to the bully or an opportunistic willingness to toady up to one. Anyone who will call it that will help you round people up or look the other way.

They're not using 1984 as a roadmap they're using Veep. www.congress.gov/bill/119th-c...