Okay, when fuzzing techniques came out there was a big surge in discovered and exploited bugs. AI is more general and I expect there be a similar surge. However fuzzing is cheap but compute and techniques can be "owned." The economics of AI is unless you pay for it, it is difficult to self host (expensive hardware, open source models are catching up).
State actors + hackers will have more resources to make better offense. What worse, in my experience AI produced code is blind to overall system behavior. So I fear the exploits will be either low hanging/trivial to exploit errors or bigger system level bugs.
>But new A.I. models like Anthropic’s Mythos, which was announced last month, appear to be so good at finding such holes that Anthropic shared it only with a limited number of firms and government agencies in the United States and Britain.
Immediate distrust of the article. GPT 5.5 is out with nearly the same capability. The author might be parroting company marketing, unable to discern that a lot of this is much less complex than it seems. For all we know this group could have had a model examine some obscure line of code thousands of times until it found something.
GPT 5.5 does not have the same capabilities as Mythos. There is a separate 5.5-Cyber model which is the Mythos “equivalent”, but it is similarly restricted access like Mythos. Per OpenAI, the major difference is the built-in safeguards that 5.5 (and other models have), where 5.5-Cyber does not have these safeguards and is more “permissive” for security work.
That is very clearly the claim of mythos though. The experience of projects that do have access to mythos though suggests that if you use the other models it's not going to find much of anything. Which is to say generally we believe it is marketing as you say however the claim that the reporter said is very clearly stated even if it's not right.
Immediate distrust of the article… The author might be parroting company marketing, unable to discern that a lot of this is much less complex than it seems.
> I am based in The Times’s Washington bureau, and much of my focus is on the dealings of U.S. cybersecurity and intelligence agencies, including the National Security Agency, Central Intelligence Agency, Cybersecurity and Infrastructure Security Agency and the Federal Bureau of Investigation, as well as their counterparts abroad, chiefly in China, Russia, Iran and North Korea.
> My remit spans nation-state hacking conflict, digital espionage, online influence operations, election meddling, government surveillance, malicious use of A.I. tools and other related topics.
> Before joining The Times, I worked at The Wall Street Journal, where I spent eight years covering cyber conflict and intelligence. My recent work at The Journal included a series of articles revealing a major Chinese intrusion of America’s telecommunications networks that breached the F.B.I.’s wiretap systems and has been described as one of the worst U.S. counterintelligence failures in history. I have also worked at Reuters and National Journal, where I began my career in Washington chronicling congressional efforts to reform surveillance practices at the N.S.A. in the wake of the 2013 Edward Snowden disclosures.
> My work has been internationally recognized, including by the White House Correspondents’ Association, the Gerald Loeb Awards, the Society of Publishers in Asia and the Society for Advancing Business Editing and Writing.
Your comment was surely well meant, but you could have plainly stated that the article author is a seasoned reporter instead of the snarky reply.
GP might be incorrect in stating that the author is parroting Anthropic's marketing, but the author certainly does not go out of his way to specify that these are only Anthropic's claims. It is actually a bit ironic as the article linked[0] from the quoted part (by another author) uses the correct phrasing when dealing with such claims:
> Anthropic, the artificial intelligence company that recently fought the Pentagon over the use of its technology, has built a new A.I. model that it claims is too powerful to be released to the public.
I feel like this website is a particularly dangerous place to ask that and hope it to be a “mic drop” moment. There are a lot of highly accomplished engineers, scientists founders CEOs, etc. here that could easily respond to that with any manner of impressive qualifications.
nytimes reporters have recently been very disappoiting and starting to feel like they're people who managed to become relevant long time ago, but haven't kept up with recent changes and are just parroting things others have said instead of unique thoughts.
OP posited that the author didn't know what he's talking about. I pointed out that the author has far more knowledge and experience in the field than rando internet griefers on HN who immediately reach for "shoot the messenger" when they read something that doesn't neatly fit into their pre-conceived worldview, instead of perhaps learning things from other people.
But at least your trope acknowledges that he's an authority on the subject.
Security will be a wedge to restrict the sophistication of open-weight and local LLMs, just as it's been used to demonize and restrict cypherpunk technologies.
> Security will be a wedge to restrict the sophistication of open-weight and local LLMs, just as it's been used to demonize and restrict cypherpunk technologies
Unlikely in America or China. This is not a game either can singularly control, and locking down the R&D means conceding momentum to the party that doesn't. Which means use restrictions will be contained to countries satisfied with playing second fiddle.
Instead, I suspect we'll see momentum towards running software on publisher-controlled servers so the source code can be secured through obscurity. It isn't perfect. But it might be good enough to get us through this transition.
If America just banned all chinese models that would wipe out most of the open weights landscape in AI, especially anything close to the frontier. I could easily see that happening if a Mythos tier model comes out of a Chinese lab in early 2027. It doesn't meaningfully change the research competition between OAI/Anthropic/Google/SpaceX but it does pad all of their pockets by removing cheap competition and it gives the government far greater control over AI usage de facto.
netsplit, I guess. decide that the risk of an open network is too great and simply block all routing out of the country through the ISPs and consider the political power that goes along with a global satellite constellation under rule of a single, government-aligned corporation.
"simply block all routing out of the country" is doing a lot of heavy lifting. For government networks, sure. For civilian networks? It's a bit like stopping pirates from ripping video; how do you deal with an attacker that ultimately can gain some form of access? Even in North Korea external media can be smuggled in.
If they tried to lock down local models more people would use them. They would also have to take down a few us companies in the process who would go down fighting for certain.
Black hat hacking seems to be a well-fit use case for these LLMs. Attackers only need to be right once, so the sometimes-wrongness of the attacks might be trivial. This probably devalues stashes of zero-day exploits for those that have been witholding them.
I wonder if that means we're going to see an increase in the attempted 'leveraging' of hoarded zero days lest they get publicised and patched prior to being profitable.
Then go ask some ChineseGPT about this, I guess, as these models seem to be much less restricted on such topics (you could even get some explosives recipes, though not all of them are real and safe) /j
Wild that they think restricting access to models will help much. Access to Chinese models will definitely not be restricted and have enough capability to find exploits as well.
The Google Threat Intelligence Group wants to increase its relevance and casually point out the it was not Mythos which found the exploit!
Security "researchers" are overpaid buffoons who hype things for their own salaries and their companies. And the stenographers from the press dutifully copy everything.
This is a despicable game to fool politicians into giving money and favorable AI legislation.
Strangely enough these buffoons never offer their models to open source developers. It is always a select group of highly paid other buffoons that throws some very occasional results over the wall.
Software is in such a state now, Gmail is full of bugs around sharing attachments to the position that I have to tell my dad to turn his phone off and on again in order to attach a document
Okay, when fuzzing techniques came out there was a big surge in discovered and exploited bugs. AI is more general and I expect there be a similar surge. However fuzzing is cheap but compute and techniques can be "owned." The economics of AI is unless you pay for it, it is difficult to self host (expensive hardware, open source models are catching up).
State actors + hackers will have more resources to make better offense. What worse, in my experience AI produced code is blind to overall system behavior. So I fear the exploits will be either low hanging/trivial to exploit errors or bigger system level bugs.
>But new A.I. models like Anthropic’s Mythos, which was announced last month, appear to be so good at finding such holes that Anthropic shared it only with a limited number of firms and government agencies in the United States and Britain.
Immediate distrust of the article. GPT 5.5 is out with nearly the same capability. The author might be parroting company marketing, unable to discern that a lot of this is much less complex than it seems. For all we know this group could have had a model examine some obscure line of code thousands of times until it found something.
GPT 5.5 does not have the same capabilities as Mythos. There is a separate 5.5-Cyber model which is the Mythos “equivalent”, but it is similarly restricted access like Mythos. Per OpenAI, the major difference is the built-in safeguards that 5.5 (and other models have), where 5.5-Cyber does not have these safeguards and is more “permissive” for security work.
See https://openai.com/index/gpt-5-5-with-trusted-access-for-cyb...
I have access to the Cyber version. It’s great at cybersecurity work but only marginally better than its predecessor with the right jailbreaking.
I imagine Mythos is going to be the same story from what I’ve seen so far.
That is very clearly the claim of mythos though. The experience of projects that do have access to mythos though suggests that if you use the other models it's not going to find much of anything. Which is to say generally we believe it is marketing as you say however the claim that the reporter said is very clearly stated even if it's not right.
Immediate distrust of the article… The author might be parroting company marketing, unable to discern that a lot of this is much less complex than it seems.
https://www.nytimes.com/by/dustin-volz
> I am based in The Times’s Washington bureau, and much of my focus is on the dealings of U.S. cybersecurity and intelligence agencies, including the National Security Agency, Central Intelligence Agency, Cybersecurity and Infrastructure Security Agency and the Federal Bureau of Investigation, as well as their counterparts abroad, chiefly in China, Russia, Iran and North Korea.
> My remit spans nation-state hacking conflict, digital espionage, online influence operations, election meddling, government surveillance, malicious use of A.I. tools and other related topics.
> Before joining The Times, I worked at The Wall Street Journal, where I spent eight years covering cyber conflict and intelligence. My recent work at The Journal included a series of articles revealing a major Chinese intrusion of America’s telecommunications networks that breached the F.B.I.’s wiretap systems and has been described as one of the worst U.S. counterintelligence failures in history. I have also worked at Reuters and National Journal, where I began my career in Washington chronicling congressional efforts to reform surveillance practices at the N.S.A. in the wake of the 2013 Edward Snowden disclosures.
> My work has been internationally recognized, including by the White House Correspondents’ Association, the Gerald Loeb Awards, the Society of Publishers in Asia and the Society for Advancing Business Editing and Writing.
What have you done lately?
Your comment was surely well meant, but you could have plainly stated that the article author is a seasoned reporter instead of the snarky reply.
GP might be incorrect in stating that the author is parroting Anthropic's marketing, but the author certainly does not go out of his way to specify that these are only Anthropic's claims. It is actually a bit ironic as the article linked[0] from the quoted part (by another author) uses the correct phrasing when dealing with such claims:
> Anthropic, the artificial intelligence company that recently fought the Pentagon over the use of its technology, has built a new A.I. model that it claims is too powerful to be released to the public.
[0] https://archive.ph/GC6WP#selection-4713.0-4713.200
> What have you done lately?
I feel like this website is a particularly dangerous place to ask that and hope it to be a “mic drop” moment. There are a lot of highly accomplished engineers, scientists founders CEOs, etc. here that could easily respond to that with any manner of impressive qualifications.
Okay, well I’ve done more than that and I say he’s right. Now what?
Reporting on such stuff requires networking skills, not technical knowledge.
Reporting on such stuff requires networking skills, not technical knowledge.
Guess how I know you've never been a reporter.
nytimes reporters have recently been very disappoiting and starting to feel like they're people who managed to become relevant long time ago, but haven't kept up with recent changes and are just parroting things others have said instead of unique thoughts.
https://www.logicallyfallacious.com/logicalfallacies/Appeal-...
Not at all.
OP posited that the author didn't know what he's talking about. I pointed out that the author has far more knowledge and experience in the field than rando internet griefers on HN who immediately reach for "shoot the messenger" when they read something that doesn't neatly fit into their pre-conceived worldview, instead of perhaps learning things from other people.
But at least your trope acknowledges that he's an authority on the subject.
> OP posited that the author didn't know what he's talking about.
That position does not appear to be present.
Eh, "unable to discern" seems like a polite way of saying someone is talking out of their ass.
How many zeroday vulns had the article author discovered using AI assisted methods?
Security will be a wedge to restrict the sophistication of open-weight and local LLMs, just as it's been used to demonize and restrict cypherpunk technologies.
> Security will be a wedge to restrict the sophistication of open-weight and local LLMs, just as it's been used to demonize and restrict cypherpunk technologies
Unlikely in America or China. This is not a game either can singularly control, and locking down the R&D means conceding momentum to the party that doesn't. Which means use restrictions will be contained to countries satisfied with playing second fiddle.
Instead, I suspect we'll see momentum towards running software on publisher-controlled servers so the source code can be secured through obscurity. It isn't perfect. But it might be good enough to get us through this transition.
If America just banned all chinese models that would wipe out most of the open weights landscape in AI, especially anything close to the frontier. I could easily see that happening if a Mythos tier model comes out of a Chinese lab in early 2027. It doesn't meaningfully change the research competition between OAI/Anthropic/Google/SpaceX but it does pad all of their pockets by removing cheap competition and it gives the government far greater control over AI usage de facto.
As long as it is within the country, restriction works. How do you restrict the capability from a foreign entity, especially a hostile one?
netsplit, I guess. decide that the risk of an open network is too great and simply block all routing out of the country through the ISPs and consider the political power that goes along with a global satellite constellation under rule of a single, government-aligned corporation.
"simply block all routing out of the country" is doing a lot of heavy lifting. For government networks, sure. For civilian networks? It's a bit like stopping pirates from ripping video; how do you deal with an attacker that ultimately can gain some form of access? Even in North Korea external media can be smuggled in.
That works for very oppressive countries. However, more freedom-minded countries are not going to law for that.
If they tried to lock down local models more people would use them. They would also have to take down a few us companies in the process who would go down fighting for certain.
Didnt work out so well with the cypherpunk technology so there is hope
Black hat hacking seems to be a well-fit use case for these LLMs. Attackers only need to be right once, so the sometimes-wrongness of the attacks might be trivial. This probably devalues stashes of zero-day exploits for those that have been witholding them.
I wonder if that means we're going to see an increase in the attempted 'leveraging' of hoarded zero days lest they get publicised and patched prior to being profitable.
Meanwhile, I cannot ask ChatGTP how to pick my own lock. Even though this information is available in a book in the library.
Then go ask some ChineseGPT about this, I guess, as these models seem to be much less restricted on such topics (you could even get some explosives recipes, though not all of them are real and safe) /j
Wild that they think restricting access to models will help much. Access to Chinese models will definitely not be restricted and have enough capability to find exploits as well.
Dupe: https://news.ycombinator.com/item?id=48096712
This is 3 hours earlier than what you're sharing.
Can we link to the actual google article, instead of these editorialized articles about the article?
https://cloud.google.com/blog/topics/threat-intelligence/ai-...
People used LLMs to find flaws in Google software.
But did they use Gemini?
I don't know, but given how often Gemini refuses benign requests IME, I would suspect it's a complete non-starter for finding security holes.
> the company added that it did not believe it was its own Gemini chatbot.
-TFA
Drives me nuts that the NYT just uncritically cites Anthropic’s unverified claims of “thousands of zero-days” without a hint of skepticism.
If "bad guy AI" can find flaws, can "good guy AI" patch them faster when backed by trillion dollar companies?
Do your AI patches introduce fewer flaws than they repair?
If I sell weapons to both sides of a conflict, can I become rich?
Yes.
Please refer to any seller of weapons ever.
No. To become really rich you have to draw a 3rd player into the conflict, and then sell weapons to them as well.
Ask anyone selling AI hardware recently!
The bottleneck is probably validating and deploying the fix, which requires coordination.
Wait until the bio version of this shows up.
The Google Threat Intelligence Group wants to increase its relevance and casually point out the it was not Mythos which found the exploit!
Security "researchers" are overpaid buffoons who hype things for their own salaries and their companies. And the stenographers from the press dutifully copy everything.
This is a despicable game to fool politicians into giving money and favorable AI legislation.
Strangely enough these buffoons never offer their models to open source developers. It is always a select group of highly paid other buffoons that throws some very occasional results over the wall.
...says yet another company hell bent on integrating it into every facet of our lives. This reads like a celebration, if you ask me.
Can google please use AI to find bugs then?
Software is in such a state now, Gmail is full of bugs around sharing attachments to the position that I have to tell my dad to turn his phone off and on again in order to attach a document
https://secgemini.google/
https://projectzero.google/2024/10/from-naptime-to-big-sleep...
https://deepmind.google/blog/introducing-codemender-an-ai-ag...
Those are all for security vulnerabilities, OP is talking about bugs with functionality.
It's probably the AI overuse introducing many of those bugs in the first place...
I can’t help but think that, Apple is big on AI and their software seems to be going to hell too.
But in exchange we get to also waste vast energy and carbon while depleting job prospects for just about any college grad.
It's not all bad though. We also managed to turn the Information Superhighway of the 1990s into the Slop Wasteland of the 2020s.