coding agents are not compilers from English to programs. and it is not because they are non-deterministic. gist.github.com/yoavg/b2454c...

it's been a really long week, and it takes a lot at this point for cyber stuff to get me to say "holy shit" even at the best of times but holy shit github.com/Bin4ry/yarbo...

The existence of a weird proxy economy for AI tokens is very effing cyberpunk, AI issues notwithstanding (or perhaps especially). (Also, China Talk is an *excellent* source for lots of current tech-related goings-on.) www.chinatalk.media/p/how-to-buy...

It’s funny because it’s true

May the forest be with you. 🌲 This May 4th, join the rebellion for clean air, protected wild places, and a planet worth fighting for. The force is strongest in nature.

Every so often I asked how to write a good conference submission. At this point, if you complete your submission without just slinging AI slop directly into the form, you're already probably in the top 50%. Being on topic, precise about novelty / prior work, and detailed in evals is top 5% material.

Seeing the top 15 tv shows meme circulating reminded me of Person of Interest. Once that show found its legs it was *so* good and ahead of its time.

Possibly the single most important statistic for any AI/ML powered security tool is the false positive rate. If you cannot tune the detector to have better than chance detection at a near-zero FPR, alert fatigue sets in and the tool is worse than useless. Base rate >>> all.

I spoke to the BBC about Mythos. When questioned about its false positive rate "Anthropic didn't mention it and sidestepped the question." High FPs is a bottleneck for defenders, sifting through results to find actual vulns may offset AI advantages. www.bbc.com/future/artic...

This article helped nail something down for me: one thing people really hate is when their experience, perspective, and ability to make decisions under ambiguity is reduced to "you are another database that needs to be exported." Treating their job as recitative and not productive.

Not enough of you folks carving your life and experience into tidy little machine-readable pieces just so you can "use AI" to optimize your life to somebody else's metric have read Seeing Like a State, and it's starting to show.

These are the kind of concerns that people should worry about rather than the unsubstantiated vulnerability armageddon. The ability to phish at this scale has been posing serious issues for a few years now.

I like the analogy of AI assistants as power tools. In the hands of a skilled professional, they're a potent force multiplier. In the hands of an amateur, they're dangerous. Both are acute effects, often immediately obvious in either direction.

"LLMs are actually pretty decent for fuzzy natural language search" and other apparently scorching hot takes.

Nobody wants to admit this but frontier models have been very good at individual vuln discovery for over a year now. The hard part has always been separating real findings from nonsense then chaining vulns to do something useful. Deterministic validation is still the foundation.

Learning to be a critical consumer of LLM output is a highly transferrable skill.

NEW RESEARCH: Yesterday, our team at @domaintools.bsky.social Investigations published a holistic analysis of DPRK malware trajectories, providing a cross-connected understanding of the adversarial landscape in Pyongyang. #infosec #cybersecurity #threatintel dti.domaintools.com/research/dpr...

All of this. Showing how models can *perform* interiority via mech interp tools is a very,very long way from being able to claim they have phenomenal experience. Particularly when purely mechanical explanations (training data) so far suffice.

This was my favourite

I don't know if this is 'hot take' territory, but: The thing that separates useful applications of AI from bullshit token churn is the ability to test the output deterministically. The more precisely you can define 'success', the better off you are. See, e.g., vuln research.

shadow IT LLM usage is never a good idea, if you allow it, it means that confidential information from your company is going into the *personal* accounts of your employees, and is subject to discovery if and when litigation eventuates. don't mix the context windows. just don't.

There are not "math people" and "not math people." School math represents a sliver of mathematics as a discipline, and there is a lot more in there that the so called "not math people" could really dig and be good at.

Man, remember data science? Remember statistics and math? Remember being able to crunch a lot of numbers through a pipeline you built yourself, that got you an analysis where you understood and could explain every step of the process and how they impacted the result? I miss that sometimes.

In earlier eras of computing unreliable hardware forced system design discipline (checksums, retries, modularity). There is a lot of that we are rediscovering in the "Agentic Engineering" era! Wrote a small post about this idea. davidgasquez.com/reliable-unr...

LLM Neuroanatomy: How I Topped the LLM Leaderboard Without Changing a Single Weight by David Noel Ng "I didn’t train a new model. I didn’t merge weights. I didn’t run a single step of gradient descent. What I did was much weirder:

For people who are concerned about having their devices seized at US airports starting Monday when ICE "assists" the TSA, EFF has this guide: www.eff.org/deeplinks/20...

Agent capabilities = agent risk. There is no real way around this yet. Now imagine an agent that's assumed huge chunks of your virtual identity and permissions getting prompt injected. We're still figuring out what to do wrt the policy controls that might let us balance the capabilty/risk tradeoff.

SwiftOnSecurity, 2014

It's very funny watching everyone slowly reinvent bits and pieces QubesOS as a response to the growth of agents and their associated risks.

We strayed from the light of god when we stopped calling it 'yolo mode'.

New Attack Against Wi-Fi

Regardless of whether Claude is 'conscious' or not (it's not, at least in any human sense of the word), you should be mindful of what habits of behavior you develop when interacting with it (or genAI in general). Practicing being shitty to something is bad for you, even if there's no Claude to harm.

behold this Pinewood Derby car that *is* a Pinewood Derby.

The only feature of 1Password that matters is their business dies overnight if they get hacked so they’ve thought harder about security than anyone you know. You can’t vibe code that in two evenings no matter how much you ask Claude to “make it secure”

a learned co-conspirator just perfectly phrased it thus: "the horrors of giving the angry vibrating crystals agency in an adversarial environment" grith.ai/blog/clineje...

personally I find combinatorics annoying so I am more than pleased that we can just let claude have at it

"Bash is all you need" -- if you give a reasonably close-to-frontier model a bash tool along with reminders about how to use sed, grep, curl, and heredocs, and remind it that it can write its own scripts and tools, you're very close to done. These things are really, *really* good at computer use.

the incentives here are so obscenely terrible that the entire thing should be illegal by default “I am sabotaging negotiations because I stand to make piles of money on Polymarket” is absolutely in play, what a horrifying clusterfuck

In case you were wondering, Polymarket had yet another spate of likely inside traders betting that the US would strike Iran by February 28. Per the due diligence investigation service Bubblemaps, the wallets used were created 24 hours earlier. The Pentagon Pizza Index has been replaced.

Saw folks fired for much less during my ARL days. Walk into a TS closed area with a fitbit? Out you go.

Well that's fun. link.springer.com/article/10.1...

Sure, giving your AI agents access to the Lethal Trifecta is an immediate broad attack surface for your life. But it also lets them do funny stuff. So who's to say if it's good or bad

The Black Hat USA call for papers is open. This will be our 6th year of having a dedicated AI track. If you have some interesting AI research, be it attacking, defending, or applying AI, we’d love to see it. Please let me know if you have any questions. blackhat.com/call-for-pap...

Agent capabilities = agent risk. There is no real way around this yet. Now imagine an agent that's assumed huge chunks of your virtual identity and permissions getting prompt injected. We're still figuring out what to do wrt the policy controls that might let us balance the capabilty/risk tradeoff.

The combination of "reasonably good guided random search" and "deterministic verifier" is both more common than you might expect and likely to remain valuable for the foreseeable. Yes, this is a AI post.

If you like the results you get from AI-generated images, or they fit your business need, god bless, but I think -- from firsthand experience over the past 2-3 years -- that if you're approaching art as a creative pursuit then you might be missing out if you use *nothing* but AI. 1/

Practical Security Guidance for Sandboxing Agentic Workflows and Managing Execution Risk from @rich.harang.org developer.nvidia.com/blog/practic...