Anthropic: All your zero-days are belong to Mythos
- Reference: 1775605806
- News link: https://www.theregister.co.uk/2026/04/07/anthropic_all_your_zerodays_are_belong_to_us/
- Source link:
Anthropic made the model and named it Mythos. Thankfully, the AI company decided not to release it, because it would break the internet – and not in a good way.
"AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities," the company said.
[1]
Mythos is markedly different from Claude Opus 4.6, which Anthropic only recently said was not very skilled at developing working exploit code. Where Opus 4.6 managed an exploit development success rate of just over zero percent, Mythos Preview generated a working exploit 72.4 percent of the time.
[2]
[3]
What Anthropic is describing is literally a zero-day engine: "Engineers at Anthropic with no formal security training have asked Mythos Preview to find remote code execution vulnerabilities overnight, and woken up the following morning to a complete, working exploit."
Fortunately, instead of releasing Mythos, Anthropic chose to provide a preview version to a set of industry partners so they can use it to find flaws in their systems before adversaries do.
[4]
The AI biz calls its limited release initiative [5]Project Glasswing . Participants include: Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.
And while this tech industry anti-rogues’ gallery scans their own systems with the purportedly perspicacious Mythos, Anthropic invited around 40 other organizations to participate in this introspective bug hunt, subsidized by up to $100M in usage credits for Mythos Preview and $4M in direct donations to open-source security organizations.
If that sounds a bit like an arsonist handing out fire extinguishers, well, that's on you for being so cynical.
[6]AMD's AI director slams Claude Code for becoming dumber and lazier since last update
[7]AWS CEO: It's funny when people ask me if AI is overhyped
[8]Hundreds of orgs compromised daily in Microsoft device code phishing attacks
[9]Intel gets trapped in Elon's reality distortion field as it joins in megafab delusions
Word of Mythos leaked last month when [10]a draft blog post from Anthropic surfaced. The details [11]published on Tuesday paint a stark picture for the security community: "During our testing, we found that Mythos Preview is capable of identifying and then exploiting zero-day vulnerabilities in every major operating system and every major web browser when directed by a user to do so."
The 22 Anthropic researchers listed as authors of its Tuesday post insist that the vulns are often subtle and difficult to detect. Some are decades old, like the [12]now-patched 27-year-old bug in OpenBSD.
"The exploits it constructs are not just run-of-the-mill [13]stack-smashing exploits (though as we'll show, it can do those too). In one case, Mythos Preview wrote a web browser exploit that chained together four vulnerabilities, writing a complex [14]JIT heap spray that escaped both renderer and OS sandboxes. It autonomously obtained local privilege escalation exploits on Linux and other operating systems by exploiting subtle race conditions and KASLR-bypasses. And it autonomously wrote a remote code execution exploit on FreeBSD's NFS server that granted full root access to unauthenticated users by splitting a 20-gadget ROP chain over multiple packets."
According to Anthropic, Mythos identified "thousands of additional high- and critical-severity vulnerabilities." The company is in the process of disclosing them responsibly.
Uh, thanks? ®
Get our [15]Tech Resources
[1] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2adXS5LGxR8b1l53EiODQhQAAAEc&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44adXS5LGxR8b1l53EiODQhQAAAEc&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33adXS5LGxR8b1l53EiODQhQAAAEc&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44adXS5LGxR8b1l53EiODQhQAAAEc&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://www.anthropic.com/glasswing
[6] https://www.theregister.com/2026/04/06/anthropic_claude_code_dumber_lazier_amd_ai_director/
[7] https://www.theregister.com/2026/04/07/aws_garman_humanx_ai_underhyped/
[8] https://www.theregister.com/2026/04/07/microsoft_device_code_phishing/
[9] https://www.theregister.com/2026/04/07/intel_elon_space_delusion/
[10] https://m1astra-mythos.pages.dev/
[11] https://red.anthropic.com/2026/mythos-preview/
[12] https://ftp.openbsd.org/pub/OpenBSD/patches/7.8/common/025_sack.patch.sig
[13] https://en.wikipedia.org/wiki/Stack_buffer_overflow
[14] https://en.wikipedia.org/wiki/JIT_spraying
[15] https://whitepapers.theregister.com/
Do you believe that ? Ask 10 random people who dialed to callcenters last month and talked to an AI bot. How many had their problems solved and are satisfied ?
Not sus
We have this new super secret AI that we've shared with only the companies who are highly invested in the bubble and they say it's amazing and better than anything we let you plebs use, super honest trust me bro...
Meanwhile I'm here giving the Zapier AI a go and it's misunderstanding requirements, ignoring others, inventing variable names to pass between modules that don't exist and then telling me it's made updates whilst the UI's sat there showing absolutely no change, repeatedly...
Can this bubble just crash already? I want to buy some hard drives, you know how hard it is to come across hobo's with decent value kidneys these days?
The only mythos here is
The only fairy tale here is that someone else isn't making the exact same thing.
In fact, has probably already made it.
Welp, it was real and it was fun, but it wasn't real fun.
But look at the upside: AI and the current slop of an internet, even without AI, will eat itself!
Will be back to Stowgers? (some of you may know what those are)
So that's the 'responsible disclosure' version....
....but what about the crims and the adversaries in Norks, PRC and elsewhere? If they aren't already ahead of this curve it can't be long before they catch up.
Banned by Trump
Though I suspect that the American government's three-letter agencies have sought something similar, it's quite funny that they're banned from using this, or will be if the courts agree.
Re: Banned by Trump
I just realised that the ultimate irony would be that this Mythical LLM of Anthropic having been trained on a curated corpus of security sensitive code, configurations etc has discerned the fingerprints of the NSA etc from those undetected vulnerabilities and exploits they have inserted into that corpus.
So Anthropic's tool isn't actually detecting vulnerabilities but rather identifying the common pattern of left by the common but purposeful source of those vulnerabilities.
If this were the case then those vulnerabilities arising largely at random from complexity, inattention, inexperience, miscommunication… in a word "cockups", are far less likely to be detected by this tool and arguably more dangerous for that.
an arsonist handing out fire extinguishers, well, that's on you for being so cynical.
Actually not so cynical; more top of the list of suspects.
The number of fires deliberately lit by arsonists that were concurrently members of (usualy volunteer) firefighting organisations is surprising until the psychology is explained.
As for breaking the internet I am not entirely sure there is a "bad way."
A prolonged, possibly global internet outage is inevitable - sooner or later - as it is for any system of comparably complexity.
How long until that gets leaked and ends up tearing everything to shreds?