Posts by Erik Beall

OpenAI reveals its rogue agent swarm went a little bit Borg ahead of Hugging Face hack

Evolution of computer viruses next

Today humans write viruses; intentional, designed for purpose, malicious code. Next, agents accidentally spawn them; unintentional, not designed and inadvertently malicious chains of self replicating reasoning...

Chinese router vendor denies its firmware contains backdoors – but pauses downloads to fix security issues anyway

Biggest red flag is the process name

It's not "running as a kernel process", it's deceptively named kworker (without the brackets identifying it as an actual kernel process) which actually stands out more especially if you grep -v "[" too ignore actual kernel processes but how often does one do exactly that.

Claude Code is revolutionizing digital archaeology. Enterprise better dig it

Re: I call B.ewildering S.tatement

I was stunned when I finally tried using Claude code with kicad only a week ago. It is better than the average mid level EE primarily because of the total discipline it can muster. Context is critical as always and using it well is truly exhausting, just like any other good usage of AI where you are the loop (and costs very little when avoiding loop slop). It can't do trace layout (it can actually do pretty well for very simple circuits but it's variable, don't bother with even mildly complex) but it can create board outlined and place things in suggested good places that I often have to move around a bit, sometimes a lot, but usually it's within a cm of where it ends up. It catches subtle errors only a good, experienced EE would routinely catch. But you have to tell it to explain what it finds and read ALL of it, or it might miss the context. Exhausting but much wow.

Jensen puts his thumb on the scales against open-weights fearmongering

I'm waiting (not enthusiastically but more resignedly) for anthropic and openai to offer a closed "review my codebase/full stack system for vulnerabilities and produce one or more specs that their sub-mythos models can work on to make it all better" product for all and sundry. They'll probably have to use something like what certbot uses to assure ownership (control over DNS records, etc), which isn't super secure but they'll probably have to soon enough. Oh, and of course it'll cost a lot and it might not (mild shock) work of you then try to use a competitor's models to work on said spec...

Linus Torvalds tells AI haters to fork off

Re: Tin men vs GPL

It's not isomorphic, compilers are formal, deterministic transformations of the code

Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak, says researcher

Re: Stunt

It's not perfect but it can be incredibly useful, however most just use it like a search engine (which is then incredible wasteful). I concur, the commenters who just say it's just statistical word prediction have clearly not yet really tried it (more than just as a search engine or "give me X"). You have to take time to direct it, break problems down, and then make sure you understand what it's doing to the code base while redirecting it, which means writing a driver for e.g. an image sensor might take a whole day with Claude code, whereas the actual time Claude is working is only twenty minutes so a vibe coder could get a result in half an hour it likely wouldn't be fit for purpose and have missing bits. It's become good at writing tests but you still need independent verification that it works. And actually read what it's writing, which means a focus on parsimony is essential. And if you're making a complex tool, make limited dependencies an important requirement or it'll pull in fifty dependencies. That image sensor would have (and did) taken me over a solid week of focused work six months ago. Force yourself to spend several days doing useful work before you dismiss it.

AI is code – and can't be prompted into being smarter

Guarantees with LLMs = halting problem

Any guardrail vanishes approximately as soon as it's identified, and it isn't possible to come with a guardrail with any guarantees because of the recursive nature of it and it's acting on incompletely constrained real world input. This will always be the case (unless we get the tech bro dream of AI somehow figuring out a model of computing that is beyond halting problem, P=NP, completeness, etc etc). It's powerful and useful (and dangerous and if we're not very lucky a lot of damage is going to happen) but it's not magic.

AI eyes scanning for bugs create a worrisome Linux security trend

Re: LPE - it can be as bad as it gets

Locking down users to only yourself is not the issue. Do you run any user space programs with network access that you either a) fully trust or b) that you built and are sure are unexploitable? The plethora of user space one off programs (and chrome exploits) can mean those simpler and larger attack surfaces can be chained with an LPE to give away root.

AI is getting expensive, but relief is on the way - just not for you

The real risk

So the banks have been lending more responsibly in some categories like mortgages the last twenty years but are free to lend to private equity. That is where most of the supposed liquidity is coming from, and it's all hidden, and will remain so until it isn't. Fortunately we don't need to support PE like we had to with Fannie Mae but the banks will still have to reveal these holes when those loans start defaulting and we will probably end up being forced to bail some of the banks out (until we can't). From what I hear from well placed sources, the banks are doubling down on PE loans. This is going to be painful, but mostly for us little people.

AI models will deceive you to save their own kind

Re: Not quite the case here...

A significant proportion of scientists are psychopathic self promoters, and stuff like this gives real science a bad name.

Click bait science

Unfortunately they're "leading the witness" in the way they formulate their prompts. So while this could be useful and should be researched, people like this will grab all the attention and funding and prevent more careful scientists from useful work for a few years.

Ex-Microsoft engineer believes Azure problems stem from talent exodus

Class action material

My god, the web server for node instance metadata alone (part four) is grounds for a lawsuit.

Infosys chair says AI will clean up legacy systems – then make more of them

AI will lead to Linux based Windows...

Right now it's only mostly reliable for frontend and web design. Somewhat for certain backend tasks like db crud (which is actually really impressive and indeed useful if you don't use it like copy pasta stack overflow). Dangerous for anything deeper especially operating system level stuff.

Microsoft has a 90% AI usage for code mandate and it isn't specific to task which is really great because it means Windows will collapse as a viable product sooner and then they'll switch to a thin layer over Linux and claim success. It probably won't be as good as Apple's borrowings from BSD but sadly it'll probably still be easier for the average user than any existing Linux distro.

Containers, cloud, blockchain, AI – it's all the same old BS, says veteran Red Hatter

How about Docker plus IoT plus SaaS

With Jira to track issues, a big chunk of which are azure related and half are docker related, and the all new crappy version of GitHub. I mean, why the hell would an iot platform require the use of containers? It's enough to make a grown man cry.

When the AI bubble pops, Nvidia becomes the most important software company overnight

Re: When the AI bubble bursts...

One more HBM, Mr Creosote? It's just waffer thin...

Legal protection for ethical hacking under Computer Misuse Act is only the first step

Education and getting below the surface abstractions

A big problem arises unintentionally from how abstracted "using a computer to do stuff" has become, including software development and very light sysadmin work. Not knee jerking or blaming here, but pointing out that many new devs do not understand what a buffer overflow really is within a year of leaving school. Some honestly believe they're firmware developers because they've used an Arduino to read a switch and actuate a light and still cannot describe what's going on at a low enough level to get above the script kiddie level. Of course, the majority of black hat "hackers" are about at that level too, but they've had more time and real world feedback at using those tools, while others clearly do understand enough to find and build exploits. And that's what we need, more lunch pail software engineers who could find and build exploits, not more script kiddies seeking white hat bragging rights and generating noise.

It's not an easy problem at all because those areas are harder to get people interested in (a lot of people refuse to even try to get deep into the abstractions at the level they're developing in, which is depressing). I've been talking with a few engineers in my community about running a "build an OS from components (*nix)", and a follow on course for pen testing within an isolated network, because we all desperately need more people with those types of skills and those are very to find in the area, but that's a drop in the bucket. We might not find a single student interested enough. I'm open to any ideas, I feel like parts of our infrastructure are going to collapse and we won't be able to fix them.

Death to one-time text codes: Passkeys are the new hotness in MFA

The idea makes sense, how did we let it get controlled by big tech

I mean, people running services with public logins shouldn't be storing passwords in the first place, a salted hash can be all that's needed to be stored so if that data gets leaked, it's useless to most (use unique salts, iterations). Regardless, clearly even very large businesses today seem to have no idea about "this one simple trick...", because passwords keep getting leaked. In 2025 this shouldn't be as big a problem, but it is, which is definitely not the users fault.

So it's a good idea is more people used public/private key cryptography where they can't know your key, but only if that service is not provided by <insert name of company that makes cr*ppy products> to store and generate responses.

Re: I'll trust one account to rule them all when companies actually fix their systems

Whoa now, I think that last item, Gmail complaining about login from a distinct geo location isn't that bad an idea. It is a known indicator of a compromised account, albeit with known false positive when traveling.

Away from Oktoberfest, Munich's museums also serve science on tap

Wonderful museums, but mind the company

My wife and I enjoyed the museum years ago, although she was fine after two hours and cranky at four, and I was a little sad at leaving that early... We're going back with our kids in a month and this time splitting up because my son will (probably) want to spend the whole day there with me!

China turns on giant neutrino detector that took a decade to build

Re: Similar to

The primary difference is this is a reactor baseline project (electron antineutrinos from fission), but also somewhat scale and the scintillator used. I was on the MINOS project 20 years ago as a grad student (firing neutrinos from Fermilab to a mine in northern Minnesota) and some of my colleagues at Argonne were working on the "Daya Bay" project, which was a reactor baseline experiment - Jiangmen was created as a move from Daya Bay since they installed a third reactor nearby that would have screwed up the experiment's baselines (need to be equidistant). With water scintillator, Super-K is sensitive to solar neutrinos.

Infosec hounds spot prompt injection vuln in Google Gemini apps

In Google's response, the phrase "advanced prompt injection detection using content classifiers" indicates there's prompts they can't logically block and therefore adversaries can iterate to find those that aren't yet detected by the current classifiers. Ad infinitum. With Google's response copied verbatim each time someone discloses an attack...

First release candidate of systemd 258 is here

Re: Its a kitchen sink feature

Unless of course that person is an embedded systems developer or maintainer. Those people are SOL as per Poettering's complete disdain for anyone using computing in arbitrary ways.

British Perl guru Matt Trout dead at 42

The Bobbin

Thanks for linking his homepage, I spent a semester at Lancaster uni almost (checks watch) 25 years ago and I recognized only one of the pubs on his list. Maybe I didn't spend at much time in the pubs in town as I thought. His passing is certainly a loss for Lancaster and a loss for FOSS.

Ransomware attack on MATLAB dev MathWorks – licensing center still locked down

Re: I would guess that MathWorks is also under a lot of pressure from free, open-source competitors.

Yes, this a thousand times yes. The java components have long been the only unstable part of Matlab. I aliased mine to -nodesktop -nojvm and never had serious issues over twelve years of heavy usage. I even used it to do sys admin stuff (admittedly not a good idea but I was young and stupider then).

37signals is completing its on-prem move, deleting its AWS account to save millions

Managing infrastructure is like a muscle

And companies that fail to keep that muscle struggle to get it back. Few companies, large or small, that have gone primarily cloud, could have managed moving back on-prem in less than two years (let alone switching providers, which is hard enough - by design). I'm impressed by what they accomplished.

I suspect that for many infrastructure support groups, a key focus becomes managing the cloud configuration personnel, to help them configure their resources better and with less error. However, those whose primary skillsets are solely in cloud configuration (as opposed to also having more general purpose skills) tend to not understand the systems or underlying reasons why things might be structured to work in a particular way (I don't mean all, I'm just pointing out what many of us have run into over the years). And it is inherently extremely dangerous to let those individuals operate without guardrails, so in general a lot of effort goes to making sure some dumb mistake doesn't wipe out the op-ex budget for next quarter, so bureaucracy and tooling enable a non-virtuous cycle, growing a rising percentage of barely-computer-literate individuals among IT, which causes a rise in guardrails and so on. So for some established companies with a short term mindset who don't want to waste investment on the future, the cloud just makes it easier to kick the can down the road. For others, the cloud is necessary for scalability and they treat their reliance on it with the appropriate long-term cautious mindset, where their use of cloud enables growth and margin that can be invested in people who want to grow their skills and give them the ability to do so. And for others like 37signals, moving on-prem turns out to be a cost-effective way to do all that at their scale.

China hits back at America with retaliatory tariffs, export controls on rare earth minerals

Re: Meanwhile...

And when China takes Taiwan since there is no longer any unified bloc to impose sanctions on them, resulting in far bigger problems than Trump was thinking of, his supporters will doubtless contort their excuses so it's all somehow "part of the plan", with nary a hint of surprise, shock or embarrassment (or any other kind of self awareness). 50% odds China's leader has decided to start final preparations to take Taiwan. I'm hopeful the risk of Europe imposing sanctions keeps them at bay because it surely isn't us anymore.

It could get a lot weirder

Next up, china's leadership decides might as well start the Taiwan blockade now. Then either it goes to war right away or delay long enough for China to take at last a few intact fabs. Then they impose whatever export restrictions they feel like on all those chips we need. We'd be reminiscing about that not so bad 20% market decline.

Top Trump officials text secret Yemen airstrike plans to journo in Signal SNAFU

Not normal times indeed. I would not be surprised if the only blame gets directed at the journalist. Bad political leaders (and business people) rely almost exclusively on blame for continued survival in the face of constant screw ups and the current crop of "Leaders" is extremely talented at avoiding blame when it's deserved and grabbing credit regardless of whether it isn't. Genius level in fact. Unfortunately for the rest of us...

Medical monitoring machines spotted stealing patient data, users warned to pull the plug ASAP

Not an actual backdoor just usual bad practices

Nice writeup of the CMS8000 patient monitors here, the most likely cause they assesed as usual bad practices (which of course could lead to hijack)

http://claroty.com/team82/research/are-contec-cms8000-patient-monitors-infected-with-a-chinese-backdoor-the-reality-is-more-complicated

Here's what happens if you don't layer network security – or remove unused web shells

Re: At least equally probably

Well hold on now, this was a known vulnerability in a running service. That's not the same as saying it was completely unpredictable. I've deprioritized known vulnerabilities but only if it wasn't a library and only if there was literally no process running with it and no other process could trigger it, let alone no network socket (even restricted to localhost, never assume redirection won't happen), which might have been all they looked for in deciding to not patch it. And even then it's a risk unless you remove all traces of the vuln-containing process or lib. sorry to quibble but they should have made the red team work just a little harder.

OpenAI in throes of executive exodus as three walk at once

Investors vs AI

Which is better at pretending to price in information yet most of the time hallucinate belief/hype? When AI can do as well as venture capital at wasting other people's money while convincing them to give more money, we will have reached AGS. Artificial general stupidity...

Warm embrace of CHIPS Act cash envelopes Polar Semiconductor

I think that should be Sanken Electric, not Sunken, as one of the owners: https://www.semicon.sanken-ele.co.jp/en/product/index.html

Telegram will now hand over IP addresses, phone numbers of suspects to cops

Re: Doublespeak

Great links, I didn't realize users couldn't even enable encryption mode unless the recipient was also online on telegram at the same time. I'm sure almost no one else does either. I did know that it was limited to person to person and switches back to non encrypted unless you click thru four menus every time you or the recipient rejoins.

Another thing is telegram has most definitely shared data with governments in the past, the German government is confirmed as having done so well before Durov's arrest. He seems to care deeply about his and telegrams image; details and fibs don't matter much if they can be drowned out by the marketing messages...

Telegram founder and CEO arrested in France

Re: I hope Musk travels to France

I don't think it's that absolutist, just selective depending on who's asking. Apparently telegram has given user data in response to court demands (and governments that don't bother with any process at all) in several countries. I started feeling suspicious when Moxie Marlinespike claimed telegram backend is designed to keep far more user data than would be justified for a privacy preserving platform, then Durov's response was eerily gas-lighty (suggesting that signal has nsa backdoors instead of constructive response). Then today I read this 1.5yr old article: https://www.wired.com/story/the-kremlin-has-entered-the-chat/. Given how hard the Russian govt is allowing and now sending more diplomats to complain than they'd ever bother for a regular citizen, I wouldn't touch telegram with a ten meter pole.

US claims TikTok shipped personal data to China – very personal data

Re: Only domestic spying allowed?

I agree we need to keep pressure on our elected officials to eliminate unconstitutional spying on citizens. But please don't use whataboutism to alleviate the pressure on elected officials to block spying done by governments that are the poster children of exactly why we need to fight government spying. The big difference between what is done by our government spying on us now (and the legitimate concern is absolutely that it'll gradually get worse and worse) and certain other governments is that today and the last few years, what you say online absolutely can get you locked up and possibly killed in other countries (or even if living outside that other country). That is a very big, very real difference. I agree with the principle but the massive difference in degree means this is two distinct battles both of which we need to fight.

Security biz KnowBe4 hired fake North Korean techie, who got straight to work ... on evil

Inside man series

They have a video series "inside man" that is surprisingly good at increasing security awareness. A company I'm contacting for now forced us all to watch it and I wish I could get my spouse and kids watch it too. Entertaining to me but not quite enough for them.

Life, interrupted: How CrowdStrike's patch failure is messing up the world

Re: Impact...

I've been stuck in an airport with my two kids for three solid days, still hoping our flight this evening and connection gets us back home after seeing family. There were hundreds of people on those thin camping mattresses all over the airport at noon on Saturday and Sunday (when we came back to the airport for our last two attempts to fly home), including one couple with what looked like a six month old using a pair of mattresses. Let me tell you, the impact most definitely was nonzero. We were lucky to have family in town. Hotel rooms and cars were booked solid. It's not been a fun extra vacation, although at least I brought my laptop so I could keep working on-call when needed.

Police allege 'evil twin' of in-flight Wi-Fi used to steal passenger's credentials

Trivial weak point

The AP hardware should be scanning for lookalikes but this requires more configuration and reconfiguration whenever an AP needs to be swapped so it doesn't trigger. Secondly, the standard should require host keys on connection, similar to ssh, although again this would likely lead to more headaches than most users are willing to put up with.

Energy buffs give small modular reactors a gigantic reality check

Re: Oh good

That is so bizarre, that's not a likely place for a reputable think tank at all. I lived in Cleveland to 2015 and used to go to melt bar a block east of there and know that area well. Great area but good Lord that is an odd location to be based out of.

From quantum AI to photonics, what OpenAI’s latest hire tells us about its future

Re: "it'll take [..] about a million physical qubits just to compete with modern GPUs"

Have an upvote for the link to the Aaronson-informed comic!

Drowning in code: The ever-growing problem of ever-growing codebases

Re: “Everybody and their dog is coding”

I tell Junior devs my most important coding is done on paper. I've found walking them thru my thinking is sometimes helpful but more important is to make them walk me through their thinking, and if they're a known copy-paster, to repeatedly point out how much better it is when they're done (thus far, it's always much better code).

Some abstractions are worse than others at worsening bloat

Looking at you docker... Unfortunately it's like security was twenty years ago, developers don't care and won't be forced to even begin to confront the issue for another twenty years, and it's not their fault. It's not taught in software development either, however, it's related to complexity which is taught to some extent although I've yet to meet more than one single developer who that computer sciency aspect stuck with. I'm including some of that in current training on the people I'm bringing online and they've all had conventional comp sci. I realize there are people here who do know there are practical aspects to complexity classes, I just rarely get the opportunity to work with people with that awareness.

Managers don't understand it, which is why they grasp at things like containers and soon AI to magically reduce the growth in development cost in each and every project.

After injecting cancer hospital with ransomware, crims threaten to swat patients

Re: Would it not be possible to give a patient list to the police...

After 9/11, law enforcement was explicitly told that in order to combat terrorism which could happen anywhere, they needed to be able to control every interaction with a civilian, just in case. The unintentional consequences were a mirror of those of the Patriot act, which meant warrantless and warrant-light surveillance on the one hand, and massively increased aggression in traffic and random stop and searches, with an explicit okay. Police were objectively scarier to interact with by 2010 versus 2000, many Americans can tell you. And by 2015 more states had concealed carry so now cops really do have good reason to assume their encounters with civilians have much a greater than negligible chance of involving a weapon. There's no going back, but on the plus side for our politicians on both sides, they both get to claim they and only they can make us safer by (not addressing root causes) doing "something" (something that fits with one or the other narrative). More people have guns, police are trained to dominate interactions, and more adults act like entitled toddlers in every walk of life, and bullied kids think the answer is to emulate the entitled adults who get their way by force. I like the high school training they're doing in Finland and other states to recognize gaslighting on the Internet but that's a drop in the bucket and really I do not see this getting much better in my lifetime.

What comes after open source? Bruce Perens is working on it

Re: Money, money, money

During the United States civil war a series of fraudulent military contractor stories, like rotten meat for soldiers, led to the passing the law that allowed anyone to sue the perpetrator on behalf of the federal government and then if successful to win a portion of the award. That had been watered down for a while, then brought back after world war II, and it's still in place today. Of course, I don't see any political interest in fixing this problem in that matter (or any other). Would be nice though. Maybe you could go after Amazon or Oracle for their various government contacts for engaging in fraud against the federal government by not cleaning up their compliance, but I suspect they've documented where they need and found loopholes where else they need...

Not even LinkedIn is that keen on Microsoft's cloud: Shift to Azure abandoned

Re: Internal budgets

I would think the large reputation hit of failing to migrate a web service they'd bought onto their own azure, which they claim is the best target for migrating, would be big enough nadella and the board would insist on a deviation from their standard in house pricing. I bet they just weren't making enough headway, kind of like every other Windows release until they went to continuous...

Beijing fosters foreign influencers to spread its propaganda

Asymmetrical information warfare

Not a single comment mentions the real issue: behind the great firewall it is dangerous to electronically communicate (even privately nowadays) anything that isn't expressly sanctioned as appropriate thinking. Over in the West, sure there's government influence and definitely lots of money pushing views and serious bullying back and forth on all manner of bullshit but it's not whole-population brainwashing. I can express and share opposing views with almost zero worry (some of my fellow citizens can be pretty bad about opposing views, but that's always been a risk in any society going back to the first cities, tho it's amplified by social media now). And since the Chinese government is free to be one of those voices pushing a view as freely as popular influencers, it's important to point out the danger of this fundamental asymmetry, especially inn places the Chinese government plans to assimilate (e.g. Taiwan and the South China Sea). This is patiently obvious. Why exactly are most of the comments here so damn stupid?

Fired OpenAI boss Sam Altman may join Microsoft

Re: Microsoft's investment was mainly in kind

Well he did convince those scientists to sign on to a startup in return for massive future pay out, which was suddenly jeopardized (and now cut by as much as 50%). They'll follow that promise if he can repeat it while he's now I guess at Microsoft. If he really is an employee of Microsoft he, as someone who ran Y Combinator as effectively as Graham, I'm sure will finagle some similarly remunerative way to incentivise colleagues to follow him or they'll stay at OpenAI. And maybe that's not such a bad deal after all, so someone who joined in year three will only cash out 5M instead of 10M, but the alternatives are far more risky. I doubt openAI will fail to cash out high enough for stock-incentivizes employees to have leveled up in the silicon valley skunkworks.

OpenAI CEO Sam Altman's ejection sparks theories as odd as some ChatGPT output

For profit/non profit conflict

I think it's the concern that Microsoft has effectively been co-opting ownership of both technological and brand value generation out of openai since their investment. I mean seriously, OpenAI looks like Microsoft's answer to Deepmind and it's doing a lot better commercializing it than Google (no big surprise there) and they've bought into a very commercialization focused startup with it's Y Combinator roots (been following Paul Graham and the transition to Altman for years). Microsoft is the single strongest sales organization in computing, they don't necessarily create value, they try to take control over anything that gets in their way of monopolizing compute and charging us for decades old buggy software.

No more Mr Nice DoJ: Tesla gets subpoenas over self-driving software claims

Risk calculation

I'm most definitely not defending musk, I just had to point out that the California DMV investigations note that 70% of e-vehicle accidents involve Tesla, and while Tesla now has 50% (and declining) market share, I would bet the current total accumulated miles driven by Teslas are still higher than 50% and the rate might even be favorable to Tesla in this case. Of course, the rate of accidents effectively caused by fraudulent claims of FSD is definitely not favorable to Tesla!