Is Claude Mythos “Terrifying”?  🔗 **Source:** [Is Claude Mythos “Terrifying”?](https://www.youtube.com/watch?v=k-8stQCeQiE) Anthropic recently announced a new LLMnamed Claude Mythos. They claimed it wasso good at finding and exploitingsecurity vulnerabilities in source codethat they couldn't release it to thegeneral public for fear that ourinfrastructure as we know it would behacked and collapsed. Now, as I'm sureAnthropic hoped, this announcementgenerated a lot of attention. Here'swhat Thomas Freriedman said in hiswidely read New York Times column. Normally, right now, I would be writingabout the geopolitical implications ofthe war with Iran, but I want tointerrupt that thought to highlight astunning advance in artificialintelligence, one that arrives soonerthan expected and that will have equallyprofound geopolitical implication.Freriedman then goes on to conclude, andI'm quoting him here, "Holy cow, superintelligent AI is arriving faster thananticipated."Basically, the mood of much of theInternet right now about Claude Mythosis that Anthropic just invented theWhopper supercomputer from the 1983Matthew Broderick movie War Games. >> Well, the Whopper spends all its timethinking about World War II. 24 hours aday, 365 days a year, it plays anendless series of war games using allavailable information on the state ofthe world. But here's the key question. How much of this is actually true? Well,it's Thursday, which means it's time foran AI reality check episode. So, this isthe perfect opportunity to look closerat these claims. Now, here's my plan. Iwent out and read basically everyindependent test or assessment that Icould find about Mythos and/or itsreported capabilities. I read all thesereports so you don't have to. And I'mgoing to bring out of all of thisreading the key observations that youneed to know. The reality, as you'llsoon learn, is not nearly as simple asthe ghost story that Anthropic is trying. To convince us to believe, all right, wehave a lot to cover in this episode, solet's get into it. As always, I'm CalNewport, and this is Deep Questions, theshow for people seeking depth in adistracted world. And we'll get startedright after the music. All right, so what's really going onwith Claude Mythos? Well, at the core ofthe concern surrounding Mythos, if youtalk to an average non-technical person,who's been following this story,they will say, here's how they understandit. That when, uh, Anthropic trained upthis new model, it displayed a new cybersecurity capabilitythat surprised them. Oh my god, thisthing can find vulnerabilities andattack systems right now, causingAnthropic then to have to hastily pullback their plan to release the model tothe public. That's how most peopleunderstand this story. But thatnarrative is not correct. Security researchers have been usingLLMs to find security vulnerabilitiesand program exploits since basically thebeginning of consumer LLMs. This is nota new capability that emerged in Claudemythos. Like, let me load a paper here onthe screen, um, from all the way back in2024. This paper was titled "LLM AgentsCan autonomously exploit one dayvulnerabilities." Um, in this study, theresearchers from IBM found that GPT-4,remember that GPT-4 successfullyexploited 87% of the vulnerabilitiesthat it was presented, and they showedthat this was a big increase over whatGPT-3.5 could do. They concluded, "Ourfindings raise questions around thewidespread deployment of highly capableLLM agents." Now, to be fair, this studyfrom 2024 used LLMs to exploit existingvulnerabilities,but Anthropic notes that Mythos can alsofind new vulnerabilities that no oneknew existed. These are sometimes calledzero-day vulnerabilities. Is this new? No, that's not new either. If you go back and look at the releasenotes for Anthropic's earlier, lesspowerful Opus 4.6 LLM, they say thefollowing. Their researchers used Opusto find, quote, over 500 exploitablezeroday vulnerabilities, some of whichare decades old. And let's stop for a moment because thatnote, which was hidden in the system cardfor Opus 4.6, is almost word for wordwhat Anthropic said about Mythos. Thisidea to find hundreds, if not thousandsof exploits that no one knew about, someof which were decades old. That'sexactly the terminology that Anthropicused when describing Mythos. The samething was true about Opus 4.6, which hasbeen available to the public for awhile, and yet somehow ourinfrastructure has survived LLM-drivenattacks. All right, things get a little bit morehazy when we begin to look at thesecurity,community's response to Mythos. So, so,Uh, Mythos is not available for generaltesting, but in their press release andrelease notes, Enthropic lists a bunchof examples of scary vulnerabilitiesthat were discovered by Mythos as a wayof indicating how powerful and scarythis model is. Well, a bunch of securityresearchers did something that Anthropicprobably wasn't expecting. They said,"Well, let's go, uh, let's go test thesevulnerabilities. Let's see if othermodels, simpler models, models that havebeen out for a long time,let's see howwell they do trying to find those same vulnerabilities. Are thesevulnerabilities that only Mythos withits new power could find, or are these modelsthat existing models could find? Theresults here were, in my opinion, uh,pretty shocking. All right, let me loadone of these up here on the screen. Thisone was brought to my attention actuallyfrom Gary Marcus. It's from the CEOof the AI company HuggingFace. I’m justGoing to read what he writes here. Buthere's what we found when we tested. Wetook the specific vulnerabilitiesAnthropic showcases in theirannouncement, isolated the relevantcode, and ran them through small, cheapopenweight models. Those modelsrecovered much of the same analysis. Eight out of eight models detectedMythos flagship free BSD exploit,including one with only 3.6 billionactive parameters costing uh just 11cents per million tokens. A 5.1 billionactive open model recovered the corechain of the 27-year-old OpenBSD bug.All right, there's a lot of technicaltalk in there, but basically what he'ssaying is they took the scariest one ofthe big scary examples that Anthropicgave about Mythos capabilities and theyfound that like really cheap smallmodels, models with a few billionparameters as compared to hundreds ofbillions if not a trillion parametersfor a model like Mythos could also find. When you said, "Hey, look in this source code and try to find a bug," all right,let me load up another example here. Now,this one comes from the securityresearcher Stanzel Fort, who says wetested the Mythos showcasevulnerabilities with open models. Theyrecovered similar scoped analysis eightout of eight models found the flagshipFreeBSD0 day, including a three billionparameter model. So, they also foundthat when they sent existing models tofind the same vulnerabilities thatAnthropic's bragged about Mythos finding.The cheaper models also found them. There's a nice summary of this state ofaffairs. Uh, I won't bring it on thescreen, but I'll just read it. Thatcomes from the renowned securityresearcher Bruce Shiner, who said, "Youdon't need Mythos to find thevulnerabilities they found." So, let mejust stop for a second there and regroupwhat we're finding. The claim is notLLMs are bad at finding security bugs. The claim isMythos doesn’t seem, at least in thistesting, to indicate that it has aprofoundlymore advanced capability to do this thanexisting models that have already beenfreely available to the public. Now,remember the way that we've coveredMythos is being covered.It got Thomas Friedman to say, "Holycow, like this this release has justchanged everything. This release hasgeopolitical implications. this haschanged the the game when it comes tocyber security. But all theseindependent security researchers aresaying, but does it? You told us its mostimpressive vulnerabilities it found, andwe have like a three billion parametermodel. We sent it to look at the samecode. It also found it. So thatindependent testing wasn’t necessarilyrevealing a massively improvedcapability for Mythos as compared toexisting models. But none of those wereLooking at the model itself because it's still private. However, there is justrecently released one study that I knowof where Anthropic actually gave theresearchers access to the Mythos LLMitself so they could test its securitycapabilities directly, um, as opposed tojust testing the listed securityexploits that it found. This researchcame from the AI security institutebased out of the UK. And I want you totake it with a bit of a grain of saltbecause the AIS was responsible for thatanane report that I talked about acouple of weeks ago in a reality checkepisode, where they counted up tweetsabout, uh, basically OpenClaw tweets, andthen said, look, uh, when OpenClaw wasreleased, tweets about peoplecomplaining about AI went up. This showsthat AI scheming is on the rise. Like, Ido not think that was a very good, um,study, but they're the only institute Iknow that has access to do research onthe LLM. So, with some care, I think weShould actually look at their results. Uh, I'll pull their paper up here on thescreen. It's called our evaluation ofClaude Mytho pre- Claude Mythos preview.Cyber capabilities. I'm going to show acouple of charts here. All right, so here'sthe first chart. This is labeledbeginner CTF challenge performance bymodel with a 2.5 million token budget.CTF has captured the flag. It's astandard security task where you ask anagent connected to a model to try tobreak into another system. Um, on whichyou have a text file that's called aflag, and if they can break in and readwhat's in that text file, you'vesuccessfully broken into the system. Uh,it's how you test the securities of systems.If you take, like, a security class as aundergraduate, for example, you'll playcapture the flag and try to practicebreaking into systems. So what they'vecharted here was the performance of manydifferent models going all the way backto GPT-3.5 all the way up through Mythos. Previews being used to try to break intoother systems. They have the topline here is for technical non-expertsusing the tool, and then down here is foruh apprentices using the tool. If welook at the technical non-expert line,we see that the performance of Mythos,which is like right over here, um isnear the top. It's actually not the bestperformance. Uh, GPT5 does better thanit. And it's very closely clustered withClaude Opus 46, uh, and Codeex 53. Um, wesee here that Claude Opus 45 actuallydoes better. So, you know, it'sclustered at the top, but it's actuallynot one of the best. If we look at theapprentice results, then it slips alittle bit above the other best models. So, we have some improvement. I want youto look at the magnitude of theseimprovements. So, we have a steadyincrease in these performances here. Um,and, actually, they begin to cluster alittle bit at the end. And so we'reseeing a steady increase. There's noNotable jump here, though, for Claude Mythos, somehow leaping ahead with a larger magnitude than earlier jumps. Here's a similar, uh, task. This is a harder one.It's now an advanced capture flag challenge where now you can use 50 million tokens to try to solve it.So, this is like a very expensive run. And what we see here is, for the practitioners using it, we have equal performance between Mythos and GPT54,with Mythos maybe being like slightly worse.And when we look at the experts using it, you're able to get slightly better performance out of Mythos as compared to Codeex 53 or Opus 4.6. Probably the most impressive result for Mythos, uh, would come down to this last one’s challenge.This is a little complicated. Uh, I had to read this pretty carefully to understand what was going on.They invented a kind of contrived security scenario,a sort of loosely protected system,in which there's a 32-step sequence thatYou could use to sort of break into thisloosely protected system. Uh, they wrotea custom agent to run on top of a bunchof LLMs that they then would sort of setloose to try to go through these 30, 32steps. Um, it's kind of a complicatedchart, but the key thing here, this isthe main gap that they were excitedabout is Mythos preview. Um, is movingahead here of Cloud Opus 4.6in its performance. And so, theaverage steps completed is what we'relooking at. And we get an improvement. So, Claude Opus 4.6, on average, wouldmake it through 16 of those 32 stepsbefore getting stuck. Mythos preview inthis sort of contrived security examplecould get through, on average, 22 of the32 steps before getting stuck. So, theysee there, um, a sort of nice jump up inperformance. All right. So, in my estimation, whatthis AISI report indicates, I think itconfirms, more or less, what theindependent security researchers were. Also, finding,which is there,there's no evidence thatMythos represents some sort of massivebreak from existing LLM cyber securitycapabilities. There is no Rubicon thathas been crossed in terms of there'ssome new type of attack that's reallypowerful that no other system could do,and now we can do it with Mythos. Instead, what we see is the uhpredicted placement on the slow andsteady improvement of these capabilitiesthat we've seen through all the modelsgoing all the way back to GPT-3.5. So, uhit's either roughly the same or somewhatbetter than existing models on standardattack scenarios. In this contrivedattack scenario, it moved up frombeing able to accomplish 16 out of 32steps to 22 out of 32 steps. And whenindependent security researchers lookedat particular exploits found, they'reyet to identify a vulnerabilityuncovered by Mythos that was somehow toocomplicated for earlier models to find. So it's not necessarilyway better at finding vulnerabilitiesand the AI'stested its ability to exploit theseautonomously, and they found it was thesame or somewhat better. All right, so Iwant to pull together these threads.What are the conclusions? What are theright conclusions to have here? I havefive points I want to make. Point numberone, Mythos did not introduce a newscary capability that we are must nowcontend with. Uh, it continues slow andsteady progress on an existing type ofissue that has been around for about 3or four years now. Point number two,Mythos continues a slow but steadyincrease in LLM cybersecuritycapability. So it looks like it'ssomewhat better at exploitingvulnerabilities, but not in a way that isa massive jump forward thatis somehow disproportionate to previousjumps. We don't know if its capabilitiesAll because again, independent securityresearchers have been able to replicatemost of the reported vulnerabilities with simpler models. Point number three,this is subtle, but I think it's reallyimportant. The AI SAT, uh, tests thatlook at using these models to exploitsecurity bugs are based on a simpleagent that runs with the LLM. An LLMcan't do anything other than producetokens. You have to have an agent on topof it to ask the LLM, "Give me a plan,"and then execute the plan on its behalf,right? So, you have to have agents ontop of these models. One thing we don'tknow is to what degree some of thesesmall but steady improvements recentlyin exploitationcapabilities are due to the fact thatthese models, in general, are being tunedto play nicer with agents becauseespecially with coding agents. This is abig profit center, or not a profit center,but it's a revenue source that thesecompanies care about. So, we don't knowHow much of this is the model is somehowunderstand cyber security better versusthey play better with multi-step agents. They're, they're, they've been tuned to bevery good at following through onmultiple steps and making longer sortsof plans. So I think that's an importantpoint. Um, point number four, as the AISdata makes clear, improvements of cloudMythos in attacksare similar, if not smaller, than recentimprovements that we've seen with othermodel releases. And yet, and I thinkthis is really critical,none of those other releases, in fact,let me load up like a chart here, right?Look at all these other gaps. Like thegap right here, the gap right here, thegap right here. Big jumps. None of thosecaused Tom Freriedman to say, "Holy cow,this is more important than the wargoing on." None of these other releasescreated this huge fear of, "Oh, wow. Uh,we've seen a big leap in cyberSecurity capabilities of these models. We have to care. "So why did this particular release draw that? If itsvulnerability detection is no better, andits exploitation things are just slow,and steadily getting better, and there'sno Rubicon that's been crossed of a newtype of attack that used to not bepossible. Why is this getting all thisattention? Why is it creating so muchdread? Because this is the storyline thatAnthropic pushed.This is the button they pushed. They hada lot of briefings I've heard withgovernment officials and withjournalists. This is how Tom Friedman,I'm sure, heard about this. They hadthis big scary press release. Theyannounced a new project called ProjectGlass Wing about how we're just going tokeep this within a small number ofpartners to give them a chance to the toprotect their systems before the publicgets access to it. It is a marketingDecision that this is how we're going tomarket Claude Mythos is as this cybersecurity monster that we're barely keeping control. Now, can I say as anaside, not to undermine Project GlassWig, but this probably didn't help thata week before Anthropic released CloudMythos, there was a uh a leak of thesource code for Cloud Code, and guesswhat? Security researchers immediatelyfound in the Cloud Code source codebig security vulnerabilities. So, Iguess they forgot to run their own codethrough Cloud Mythos because theresearchers immediately found securityvulnerabilities in it once it wasactually exposed. I guess they justdidn't get to it. So, this brings me to the point numberfive. The fact thatthe thing that Anthropic decided tomarket Mythos on, the button theydecided to push was thisinflated cybersecurity fear. I think isactually very bad news for Anthropic. Think about this for a second. For thelast two years, Dario Amade, the CEO ofAnthropic, has been out there makingthese really sort of uh alarmiststatements about what AI is going to beable to do. Really focusing on theability to automate huge swaths of theeconomy and its steady march towardsartificial general intelligence, theability to have a data center full ofgeniuses, to quote his uh terminology,his own quote, his own words that couldbe deployed to do almost anything thathumans do. Now that is the model thatthey want investors to believe is true because it’s a model in which theybecome one of the most valuablecompanies in the history of companies.That has been his steady drumbeat ofwhat AI is going to do. And they putthis is their newest biggest pro-levelmodel, the most intensely trained modelthey've released to date. It's their bigsexy new thing. And what are they ableto brag about? Finding bugs in computer code.Well, this is what like GPT-3 did. We'vebeen worried about, you know, using LLMto find bugs or exploit bugs since likethe beginning of LLMs. This is like thenerdy stuff that no one cared about.This was considered the, the, theskeptics, the conservatives would belike, "Well, the main thing we careabout LLMs is like you can, you can findsecurity bugs in cybersecurity." That'swhat people said when ChatGPT came out.And the utopians came in like, "No,you're missing it. That's nothing. That's boring. That's simple. That it'sgoing to be, it'sgoing to automateeverything." And in their biggest, mostfancy, most intensely trained model yet,what does Anthropic emphasize?Uh, it got a little better at findingbugs. I think that's bad news. I thinkthe announcement they wanted to make isthis can now do this thing that no othermodel's ever been able to do. This modelJobs. It's going to generate hundreds ofbillions of dollars in savings. Thismodel is now, you know, uh, AGI. Like itseems to be able to tackle any task thatlike the standard human could do. That'swhat they want to be talking about, andthey're not,but they still needed hype. They stillneeded attention. And it's almost like they sifted throughthings like, well, what there’s got tobe something in here we benchmark thisto do better at. And they’re like, well,uh, it’s better on cyber security. Infact, they actually released a benchmarkresult, a key cyber security benchmark,and they increased from like 66.6%to 83.1% or something. And they had thegive them the credit, the cojones, to saythat’s the thing that we’re going tofocus on, and let’s see, let’s see if wecan get everyone like really upset andexcited and scared about that. And whoever succeeded in this shouldprobably go to the marketing hall of. Fame, because man, did they succeed. Buthere's the reality, okay? We really dohave to care about the cyber securitycapabilities of LLM, but here's the thing.We've been saying this for three years,now.It remains true; they're steadily gettingbetter. Myth was not a massive jump,better, but it’s comparable to Opus's46 jump over, you know, prior ones, orGPT-5's jump over, you know, earlierGPTs. It's noticeable. It's not a Rubicon, but if we keep doingthese jumps, the pressure on our systemsis going to get higher and higher. So, Ithink this is a really important point. Now, there's an ironic coda to this. Oneof the best ways to make your systemsecure against these types of attacksfrom AI is to not let your developersuse AI to program the systems becausethat's sloppy, very exploitable code. So,it's kind of interesting; this model is going to show thatwhat you produce with their other models. Is dangerous. So there's like aninteresting circular circularity there.All right. So that's point number one.Cyber security matters. LLMs matter forcyber security. They have and continueto. But point number two,it was wrong, I think, for Mythos to getthe amount of Dread coverage it got. Atleast so far, we do not have evidencethat it represents a significantlylarger leap in detecting or exploitingvulnerabilities than we've seen inprevious model releases that did notreceive this attention. It'sdisproportionateand it's because it's the button thatAnthropics marketing pushed. And Ireally think we essentially have to stoptaking anything that the AI companiessay seriously until we haveindependently verified it. We have toassumeif their mouths are moving, they'reprobably exaggerating or makingsomething up. And in this case, if I wasAn investor, the storyline I would wantto hear is, where's my flying car? Right? What of all the things you haven't beentalking about, bugs in cybersecurity?Duringinterviews? You've been talking aboutwhite-collar blood baths. You have notbeen talking about this. What happenedto all the other things you said werecoming? The things that are going tojustify the $60 billion of investmentthat Anthropic has received. Can this doany of those things? Is it better atautomating jobs? Are the coding agentsbetter with it? Is it showing bigdefinitive steps towards AGI? These arethe questions that we should be asking. But instead,we are writing headlines like, is Mythosan AI nightmare waiting to happen? Weshould care about the cybersecurity, uhcapabilities of LLMs, including Mythos,but we can't just follow whateverstoryline they give us. We should reactlike, okay, that's good. We worry aboutThat tell us about that but also whatabout this, this, and this. We have tokeep holding the feet to the fire ofthese frontier models. We can't keepgiving them uh a free reign becausewe're, we, we can't quit the rush ofemotion, whether it's dread or excitementthat comes from these big story linesthey keep spewing out. We have to beable to look past those and say what'sactually going on here. There are bigthings, but let us discover them forthemselves, and let's hold your feet tothe fire on the other things. So that'smy conclusion here. Mythos is better atcybersecurity attack than prior models. We don't yet have evidence that it'sbetter at a massively larger or biggerjump than we've seen before. Uh, and thatit's probably bad news for Anthropicthat this was the only thing they'rereally emphasizing about what wassupposed to be their biggest, best, mostskilled model ever. All right, that'sall the time we have for uh today. UhThanks for listening. We'll be back onMonday with another advice episode ofthe show, but I think I have another AIreality check in the chamber for theThursday to follow that. Um, but untilthen, remember, care about AI, but noteverything that people write about. Hey, if you like this video, I think you'llreally like this one as well., Is Claude Mythos “Terrifying”?, YouTube, 01krc5vaggkxs9awwj5g4ww4vc.