BSides NoVA was so fun to speak at last year. Their CFP is now open, y’all should join me! sessionize.com/bsidesnova-2...

"Happy" "There Will Come Soft Rains" Day, to all who celebrate

I usually resist dunking on/amplifying these terrible takes, but holy shit. I know some day, probably not too far off, my kid's gonna start having too much going on to want to sit and shoot the shit with me all that often, and I'll be sad when it finally happens. Why would you ever give that up?

Is it me or does each new generation of Claude models seem to clutch their pearls just a little bit tighter than the last?

You can produce a similar effect by exhausting your red cone so that it stops firing on green light. It doesn't last long but the color is very intense. dynomight.net/colors/

New colors without shooting lasers into your eyes

Can optical illusions take you outside the human color gamut?

dynomight.net

The New Yorker@newyorker.com · 6d ago

Researchers at U.C. Berkeley used lasers to stimulate retinal cone cells in a way that doesn’t occur naturally—creating a new color that only seven people in the world have seen. www.newyorker.com/magazine/202...

Latest from the NVIDIA AI Red Team (with @beccalunch.bsky.social) -- security controls that make attacking interactive agents hard. The most annoying ones were good old-fashioned ones: access control and tightly scoped network egress lists. Enjoy! developer.nvidia.com/blog/four-wa...

Four Ways to Deploy More Secure AI Agents | NVIDIA Technical Blog

Knowledge workers are increasingly integrating AI agents into their workflows. Agents that function as “digital coworkers” offer clear benefits. For example, they can review a bug report…

developer.nvidia.com

The Hugging Face incident was “both notable and alarming,” EFF’s @evacide.bsky.social told @sfexaminer.bsky.social, “not because ‘ooh, the model’s so powerful,’ but because it demonstrates a colossal failure on the part of OpenAI to secure the sandbox in which it was testing its model.”

OpenAI hacking incident highlights AI cyberrisks

Experts take company to task for safety efforts

sfexaminer.com

Anecdotal and vibes, but Opus 5 and Fable both seem distinctly worse than GPT 5.6 at multistep reasoning outside of coding. The Anthropic ones also careen wildly between rank sycophancy and getting extremely pissy when you push back. Hard to trust, sometimes annoying to use.

Reposting and quote posting for feed reach.

DEF CON@defcon.bsky.social · last wk.

Friendly #DEFCON34 reminder: Meta-style glasses with recording capabilities are prohibited at DEF CON. There is no exception for recording glasses with prescription lenses. Be sure to pack non-violating eyewear if you need them. Please see our Photo policies for more. defcon.org/html/links/d...

Adjust threat models not just for being the victim but also the attacker. New paper by many authors gives a detailed set of recommendations, supporting my initial assertions last week that orgs need to assume their own agents could attack others & factor that into agentic AI risk

Gadi Evron@gadievron.bsky.social · last wk.

Releasing: Post mortem analysis of the Hugging Face incident was written over the weekend by hundreds of CISOs (and reviewed by Hugging Face). Link: cloudsecurityalliance.org/artifacts/hu... (+free download) From CSA, SANSInstitute, Knostic, [un]prompted, RSAC, FIRST

Open source is critical infrastructure for the global economy. Launching today, the Open Secure AI Alliance brings industry and community together around shared research, tools and vulnerability harnesses to help defenders find and patch bugs before attackers strike. nvda.ws/4pAMWBy

I try not to post too much NVIDIA-specific stuff, but I'm glad NVIDIA is pushing this: blogs.nvidia.com/blog/open-se... Attackers have demononstrably abused access to frontier models to drive operations. Defenders, constrained by TOSs, don't have the same tools. We need to level the playing field.

Industry Leaders Unite in Open Secure AI Alliance for AI Safety and Security

NVIDIA and founding members form new alliance to build and share open tools that promote responsible use of and trust in AI.

blogs.nvidia.com

Incredibly frustrating to see the way big orgs try to vacillate between 'AI is a tool, the human is responsible when it does harm' and 'no, the AI is responsible' based almost entirely on what it would cost them to own it. Humans are only a moral crumple zone (a la Elish) when they're not C suite.

Eryk Salvaggio@eryk.bsky.social · 2w ago

Seeding the “maybe our model is exhibiting self-awareness?” narrative to obliterate the “OpenAI is a negligent and potentially criminal actor” narrative is a reputation management move exclusive to this industry. Not a stunt, but a desperate need to control the narrative.

Definitionally, these are not “friends.”

Warning: Your Friends and Colleagues Are Now Recording Every Conversation

People in tech circles are carefully recording every single conversation they're having — not just at work but in their private lives as well.

futurism.com

I was a 911 dispatcher for a decade. To really understand how terrible an idea it is to have AI answer 911 calls, I'll present you with a single real-life call: Upon answer, the caller calmly began ordering a pizza. Large, pepperoni. Side order of chicken wings. To be delivered promptly.

Alex Hanna@alexhanna.bsky.social · 2w ago

AI has no place in 911 calls. From Mystery AI Hype Theater 3000’s latest episode “State of Emergency”. (with @emilymbender.bsky.social)

This matters a lot because people are constantly asking me if they should set up a duress passcode in order to fool law enforcement.

Zack Whittaker@zackwhittaker.com · 2w ago

New, by me: The Justice Department is prosecuting an American for allegedly providing U.S. border agents with a "duress" passcode that wiped the contents of his phone when they entered it. We've confirmed the phone was running GrapheneOS. Bypass for ad-blockers: web.archive.org/web/20260724...

"PROMPT HARDER" remains an ineffective security control.

I keep saying the strength is in the harness, not the model - because it's true. No use without a harness, though. Here's VISA's open-source cybersecurity harness. Just add model! Very cool of them to share this tech and lift the defensive cybersec poverty line. github.com/visa/visa-vu...

GitHub - visa/visa-vulnerability-agentic-harness: Visa Vulnerability Agentic Harness

Visa Vulnerability Agentic Harness. Contribute to visa/visa-vulnerability-agentic-harness development by creating an account on GitHub.

github.com

Last comment (for now) on The Events: The fact that frontier models caught the vapors and refused to help HF triage the attack is another example of how overly fussy "safety guardrailing" is forcing defenders to fight by Queensberry Rules while attackers are busy shanking people in back alleys.

LLMs / agents / whatever do not inherently have things like code execution, bash access, network access, etc. We decided at some point that, despite the now obvious and repeatedly demonstrated risks of these tools, they should be baseline features of agentic tools. We can make better decisions.

Re-upping this. An inference-only LLM is basically inert. It only becomes dangerous when you allow it to do things like write and execute its own code, or run shell commands. Stop hooking the goddamn things up to tools they don't need.

Rich Harang@rich.harang.org · 2mo ago

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.

My book, Reinforcement Learning from Human Feedback is done! This is the book I wish I had when learning to fine-tune, align, & now post-train models since ChatGPT. The resource has been built by me finding time to study and document the fundamentals on nights and weekends since 2024.

Hey, that's a nice [THING YOU DO FOR PERSONAL PLEASURE], be a real shame if somebody were to [CONVINCE YOU TO MONETIZE IT, TRANSFORMING YOUR RELATIONSHIP INTO AN OBLIGATION AND ROBBING YOU OF JOY]

At the airport: "we discovered a maintenance issue, but don't have any maintenance staff on site during the weekend, so it'll be 45 to 60 minutes before they can be here to take a look at it" How the fuck do you not have maintenance on site whenever you're actively moving aircraft?

the original nerd snipe: ancient geometers looking at a circle like "that's gonna bother me forever, actually"

Told a coworker today that there was a time when 1.4MB was considered a useful unit of storage, and while they didn't call me a liar directly, there was enough skepticism that it was at least implied.

Search Google for MITRE ATLAS -- a terrible AI summary, four sponsored results, six suggested searches, a bunch of youtube videos, four more suggested searches, social media results (what?), and finally, an actual result. Followed by another sponsored result and six more suggested searches.

Air quality at 6:00 today, after the 5th of July fireworks that didn't start until the 4th had been over for 30 minutes. At least the incessant fucking fighter jet flyovers stopped early for weather.

Paraguay to the World Cup audience:

After the CPV ARG game this one is so disappointing.