https://www.nytimes.com/2026/04/07/opinion/anthropic-ai-claude-mythos.html?unlocked_article_code=1.ZVA.DuZ1.tSnsJb7Od3ZD&smid=nytcore-android-share

Quote:

---

Anthropic said it found critical exposures in every major operating system and Web browser, many of which run power grids, waterworks, airline reservation systems, retailing networks, military systems and hospitals all over the world.

If this A.I. tool were, indeed, to become widely available, it would mean the ability to hack any major infrastructure system a hard and expensive effort that was once essentially the province only of private-sector experts and intelligence organizations will be available to every criminal actor, terrorist organization and country, no matter how small.

---

According to the writer, Anthropic developed an AI that got so good, so quickly at finding security flaws that it "scared them" and they're no longer planning a wide release out of fear of the inevitability of misuse. Instead they reached out to the govt and other major tech companies to help ensure the weaknesses it found were shored up.

Scary stuff, and the province where many future battles will be fought I'm sure. Makes me want to be a Luddite.

Anthropic is currently on the **** list with the federal government so I wonder how much of this is really about trying to get back into the feds' good graces.

They've made a lot of claims previously, such as Claude had become AGI, which were later found to be untrue.

Quote:

---

They've made a lot of claims previously, such as Claude had become AGI, which were later found to be untrue.

---

And this is the problem with the current AI discussion. Anyone that has consistently invested in or with or dealt with VC stage tech firms by reflex raise a cynical eye and make sure their wallet hasn't been stolen after any such headline release.

But the larger general public is taking everything at face value and assuming total authenticity.

Things are going to move fast in the next few years.

I read an article this week (sorry...not going to try to find it) that quantum computers are a big enough step forward that they'll soon be able to solve bitcoin encryption faster than the transaction can be completed. So they'll be able to hijack online transactions in progress even if you keep your bitcoin offline.

One of the guests on Fox Business yesterday (forgot his name) pitched the theory that the US and China would go together to try to figure out how to defeat any AI attempts to throw the respective countries into turmoil. His belief is that a country's own citizens would be responsible for such hackery.

And yet they themselves accidently leaked some of their source code

Logos Stick said:

---

Claude is now the best software engineer and top cybersecurity expert in the world. With Mythos, I can have Claude quickly decompile and disassemble any code on my desktop and find vulnerabilities. This is a serious national security issue, imo.

---

Derive source code from a binary?

Logos Stick said:

---

Claude is now the best software engineer and top cybersecurity expert in the world. With Mythos, I can have Claude quickly decompile and disassemble any code on my desktop and find vulnerabilities. This is a serious national security issue, imo.

---

And yet, every software engineer points out that LLM's only understand coding on a small scale and have no concept of data structures, organization, ease of maintenance, etc. Left to its own devices, Claude will generate a pile of junk code that will take twice as long to fix as it would to simply start fresh.

So I have finally found an X post from a credentialled expert who has actually reviewed the thing. She notes a lot of things are missing required to judge the true power of the platform.

Dr. Heidi Khlaff said the following . .in total "here are some red flags that should give you caution in taking these claims at face value."

• No comparison benchmarks provided with traditional static analysis. tools well-used within current AI platforms might be doing just as well.
• No discussion of rate of false positives. AI security platforms are famous for mistakenly identifying secure platforms as insecure.
• No detail on the scale of human intervention in the process. They confirm internal security experts were part of the process but not how many and the scope of their involvement
• She dislikes the Anthropic framing of open source code as well audited. She says most mission critical systems have open source bans in place for as there has historically been little auditing. Anthropic is choosing an easy mark and potentially over-extrapolating the significance.

The substacks I read tend to view this as pre-IPO marketing stunts that nevertheless do highlight incremental benefits of AI powered platforms in enhancing the speed of cybersecurity.

Rapier108 said:

---

Anthropic is currently on the **** list with the federal government so I wonder how much of this is really about trying to get back into the feds' good graces.

They've made a lot of claims previously, such as Claude had become AGI, which were later found to be untrue.

---

The feds won't come begging. What Anthropic produced is at most 3-4 months ahead of the competition. All the frontier AI firms have 10T parameter models in training.

Stmichael said:

---

Logos Stick said:

---

Claude is now the best software engineer and top cybersecurity expert in the world. With Mythos, I can have Claude quickly decompile and disassemble any code on my desktop and find vulnerabilities. This is a serious national security issue, imo.

---

And yet, every software engineer points out that LLM's only understand coding on a small scale and have no concept of data structures, organization, ease of maintenance, etc. Left to its own devices, Claude will generate a pile of junk code that will take twice as long to fix as it would to simply start fresh.

---

I generally agree, but they could probably be taught, the same as a human.

Can LLMs speak up when they don't k ow something and need to be shown? Or can they only confidently spout nonsense?

Complex code is still an issue, however Claude can easily handle more simple code, especially when it comes to query languages. While your seasoned vets are still safe, Claude can accomplish in 5 minutes what you used to hand off to an intern or a junior developer. It will be interesting to see what environment the graduating class of 2030 has to deal with.

Hmmm, was the China super computer hack a test case?

Our phones are going to become useless, the sales attack ai is relentless.

The Gemma 4 model is more likely to be world changing. The number of parameters going up doesn't really help as much as the way Google just changed the way memory is used (and did it with a true open source license). This Mythos stuff is just marketing hype.

Did it actually find flaws or just tell the person searching for them what they wanted to hear?

AI, taking over the world!?



Tttttrrrrrruuuuuummmmmmmpppppp!!!!!!1

I've used Claude and several other AIs quite a bit in an attempt to find help in doing scholarly research. The positive is that they, Claude in particular, can suggest ideas that I had not even considered. Nor, as best I can tell, has anyone else ever considered them. In other words, Claude appears to have original ideas.

The bad is that the net output is worthless. Every idea, no matter how original, has to be anchored in some reality. Claude will cite articles in support of his/its novel ideas, but the articles turn out not to exist. Claude readily admits that it is hallucinating, but admits so in a very friendly, disarming manner.

It raises the question, in my mind at least, how much the output of these AIs can be completely trusted. I came across an article recently in which the author claimed that these flaws cannot be cured but are baked into the very hardware of the AIs. Is that correct? I have no idea. But his thesis is that we are quickly reaching the ceiling for the AIs, rather than the exponential improvement that many AI bros are claiming.

My personal experience, using AIs for things like scholarly research, and mundane things like shopping for the best prices, is that AI output cannot be trusted to be accurate at all.

Stmichael said:

---

Logos Stick said:

---

Claude is now the best software engineer and top cybersecurity expert in the world. With Mythos, I can have Claude quickly decompile and disassemble any code on my desktop and find vulnerabilities. This is a serious national security issue, imo.

---

And yet, every software engineer points out that LLM's only understand coding on a small scale and have no concept of data structures, organization, ease of maintenance, etc. Left to its own devices, Claude will generate a pile of junk code that will take twice as long to fix as it would to simply start fresh.

---

As the owner and architech of some enterprise AI software - let me tell you this isn't the case. Claude and ChatGPT are kings of development right now and both are good enough to be compared to an extremely fast Jr developer. Talk to it right and I can get 10x more code done than I ever could on my own.

Past couple of months I'm leaning heavy into the software testing side. Claude - design this new feature I want. Claude - build this new feature out. Claude - Test this feature. Claude - Deploy the feature and test in the qa environment. If there is any problem fix it and come back to me.

I watch Youtube for 15 minutes. Come back and everything is done.

Claude is the better creative, ChatGPT finds bugs better. Google needs to show me something before I go back there and I don't want to deal with Gemma on my laptop right now.

A new iteration of Claude could be an incredibly massive change to how software is written.

KingofHazor said:

---

I've used Claude and several other AIs quite a bit in an attempt to find help in doing scholarly research. The positive is that they, Claude in particular, can suggest ideas that I had not even considered. Nor, as best I can tell, has anyone else ever considered them. In other words, Claude appears to have original ideas.

The bad is that the net output is worthless. Every idea, no matter how original, has to be anchored in some reality. Claude will cite articles in support of his/its novel ideas, but the articles turn out not to exist. Claude readily admits that it is hallucinating, but admits so in a very friendly, disarming manner.

It raises the question, in my mind at least, how much the output of these AIs can be completely trusted. I came across an article recently in which the author claimed that these flaws cannot be cured but are baked into the very hardware of the AIs. Is that correct? I have no idea. But his thesis is that we are quickly reaching the ceiling for the AIs, rather than the exponential improvement that many AI bros are claiming.

My personal experience, using AIs for things like scholarly research, and mundane things like shopping for the best prices, is that AI output cannot be trusted to be accurate at all.

---

Hallucinations are baked in because it's really a next word predictor based on training data. It doesn't know facts from fiction. It doesn't use logic or reasoning. It's interesting that some of the ideas appear to you to be novel. Maybe because its training is on word associations, and your training is in a research field?

AI is really good at code because code is so structured. The prediction of the next word is a lot easier as a result.

It just returns words that look a lot like words in the training data.

I think that it is short-sighted to think that there is no possible fix for the hallucinations thing.

Again, one way to sniff them out is to turn the AI on itself. It gives you an argument, you ask it why that argument could potentially be wrong, and zero in on the things that it identifies.

Once you get the AI to do that well automatically, a lot of these hallucinations may go away.

Or, there may be some other way forward that we haven't even thought of yet. But to say that it is an impossible problem to fix is short sighted in my opinion.

BusterAg said:

---

I think that it is short-sighted to think that there is no possible fix for the hallucinations thing.

Again, one way to sniff them out is to turn the AI on itself. It gives you an argument, you ask it why that argument could potentially be wrong, and zero in on the things that it identifies.

Once you get the AI to do that well automatically, a lot of these hallucinations may go away.

Or, there may be some other way forward that we haven't even thought of yet. But to say that it is an impossible problem to fix is short sighted in my opinion.

---

Hallucinations a year ago vs Hallucinations today are night and day differences. It's getting so much better.

Quote:

---

Hallucinations a year ago vs Hallucinations today are night and day differences. It's getting so much better.

---

Most of the studies show that very limited reasoning has shown some improvement but more complex queries are making the problem worse.

whoops replied to wrong person!

You gotta up your prompt game. Instruct the AI to anchor responses in research and facts, and to provide citations. After that, start a new session and tell the AI to red team/fact check the output of the first.

Finally, I have to ask: are you using a paid or free version? A frontier reasoning model with extended thinking is night and day different from the free crap.

KingofHazor said:

---

I've used Claude and several other AIs quite a bit in an attempt to find help in doing scholarly research. The positive is that they, Claude in particular, can suggest ideas that I had not even considered. Nor, as best I can tell, has anyone else ever considered them. In other words, Claude appears to have original ideas.

The bad is that the net output is worthless. Every idea, no matter how original, has to be anchored in some reality. Claude will cite articles in support of his/its novel ideas, but the articles turn out not to exist. Claude readily admits that it is hallucinating, but admits so in a very friendly, disarming manner.

It raises the question, in my mind at least, how much the output of these AIs can be completely trusted. I came across an article recently in which the author claimed that these flaws cannot be cured but are baked into the very hardware of the AIs. Is that correct? I have no idea. But his thesis is that we are quickly reaching the ceiling for the AIs, rather than the exponential improvement that many AI bros are claiming.

My personal experience, using AIs for things like scholarly research, and mundane things like shopping for the best prices, is that AI output cannot be trusted to be accurate at all.

---

If you're using a chat directly with the AI, definitely not. If you're using a multi-agent system that can do validation and error checking (Agent A gets the answer, Agent B validates Agent A, and back and forth). Some multi-agent systems will go to the level of using multiple LLMs to validate data. But, because you're doing all those back and forth calls, it gets expensive fast, especially if you don't provide guardrails on how many times it'll loop.

BusterAg said:

---

I think that it is short-sighted to think that there is no possible fix for the hallucinations thing.

Again, one way to sniff them out is to turn the AI on itself. It gives you an argument, you ask it why that argument could potentially be wrong, and zero in on the things that it identifies.

Once you get the AI to do that well automatically, a lot of these hallucinations may go away.

Or, there may be some other way forward that we haven't even thought of yet. But to say that it is an impossible problem to fix is short sighted in my opinion.

---

I agree the hallucinations are overblown and can generally be avoided/minimized,

First AI running in several roles, providing its own validation. THEN a second AI checking the work of the first.

A pain to set up, but little additional work once it is set up.

Does not mean it cannot be very useful under the right circumstances and save some time in the right hands, though. It has utility.

Ultimately it is a tool. It is a technology. It has usefulness.

The problem is it's very expensive to develop and run. A lot of these companies will fold before they ever turn a profit. They have to oversell what their products can do to drive investment.

The VCs who invested in this a few years ago after Chat-GPT amazed them with its magic are pushing out narratives that are completely unrealistic, in my opinion to eventually con retail investors to hold the bag while they run for the exits and/or push public sentiment about a race to AGI to justify some kind of government bailout.

Everyone should definitely trust Claude implicitly.

Rydyn said:

---

Things are going to move fast in the next few years.

I read an article this week (sorry...not going to try to find it) that quantum computers are a big enough step forward that they'll soon be able to solve bitcoin encryption faster than the transaction can be completed. So they'll be able to hijack online transactions in progress even if you keep your bitcoin offline.

---

There is a a fair amount of controversy about this. On of the biggies was doing testing on encryption using quantum, the tests were unexpectedly stopped by the US govt. Read into it what you will.