AI Judge changed title after evaluation, original title: "Calif researchers uncover macOS vulnerabilities in Mythos AI testing"
Calif researchers discovered two previously unknown macOS kernel vulnerabilities while testing an early build of Anthropic’s Mythos AI in April. They chained the flaws with additional techniques to corrupt memory and bypass Apple’s hardware-backed protections, reaching inaccessible device regions. The privilege escalation exploit was completed in five days of evaluation. The team submitted a 55-page report on the findings to Apple.
Well, well, well... Good for hardening the system
ANTHROPIC’S MYTHOS HELPED RESEARCHERS FIND MACOS BUGS
WSJ reports Calif researchers used techniques from Anthropic’s Mythos AI testing to find two macOS bugs.
The bugs were linked into a privilege escalation exploit.
Apple is reviewing a 55-page report and details are expected
Positive users celebrate Mythos AI legitimately cracking macOS security in days as proof of real progress from Anthropic, while negative users dismiss the reports as exaggerated marketing theater or underhanded fear-mongering.
No Digg Deeper questions have been answered for this story yet.
Mythos has cracked MacOS. It took five days.
The Second Scaling Law remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just add thinking tokens. There doesn't seem to be any plateau so far.
Very important update from UK AISI. This is a meaningful change from the previous report. Here’s what the new data would look like for “Mythos Preview (new)” with $ on the x-axis:
Very important update from UK AISI. This is a meaningful change from the previous report. Here’s what the new data would look like for “Mythos Preview (new)” with $ on the x-axis:
Our cyber range results illustrate this step-up. Since our first Mythos evaluation, we received access to a newer Mythos Preview checkpoint. On a 32-step corporate network attack we estimate takes a human expert ~20 hours, this checkpoint completes the full attack in 6 /10 attempts.
WSJ: Anthropic’s Mythos helped researchers find 2 unknown macOS kernel bugs and turn them into a working privilege escalation exploit in 5 days.
The target was the macOS kernel, the deepest layer of Apple’s desktop operating system, where code controls memory, processes, permissions, and access to hardware.
Mythos helped connect 2 separate flaws with extra exploitation techniques, which means the attack did not rely on one bug but on a chain where each step made the next step possible.
The exploit allegedly corrupted memory, bypassed Apple’s memory integrity protections, and gained access to protected parts of the system that normal apps should never reach.
This is serious because modern macOS defenses are built to make memory bugs hard to convert into control of the machine, not just hard to find.
Mythos can become so powerful here because vulnerability research is a search problem with many dead ends, where the model can help form hypotheses, inspect code behavior, reason across low-level constraints, and suggest exploit paths faster than manual work alone.
---
wsj .com/tech/ai/anthropic-mythos-apple-macos-bug-339da403
Head of Frontier Red-Team at Anthropic.
A lot of people have been wondering about Mythos, Glasswing, and the vulns we / our partners are fixing. Today, I’m excited for us to start sharing more. (For context, I lead Glasswing @AnthropicAI.)
Two independent evaluations this week—from XBOW and the UK AISI—confirm what we've been seeing internally: Claude Mythos Preview is a step change in autonomous cybersecurity capabilities. We need to start preparing fast for a world of models with this level of capabilities.
The UK AI Security Institute tested the model we shipped at the launch of Project Glasswing and found Mythos Preview is the first model to solve both of their end-to-end cyber ranges, including one (Cooling Tower) which no model had ever cleared. But attackers (and defenders) have sophistication & cost constraints – Mythos is also the only model that clears every one of their tasks estimated over 8 hours under their deliberately low 2.5M-token cap.
XBOW tested it on their offensive security benchmarks, finding "token-for-token, unprecedented precision." It's the only model to succeed at subtle V8 sandbox work.
Other Glasswing partners shared similar stories. In a few weeks of testing, Mythos Preview has helped them find many thousands of (estimated) high + critical severity vulnerabilities, sometimes double what they'd normally find in a year.
I don't share this to boost Mythos. In fact, this is not about Mythos. It’s about preparing for the coming world of models being better, faster, cheaper, and more creative than some of the best human experts at dual use capabilities. Clearly, we need them supporting defenders as widely as can be done safely – and especially the least resourced ones.
Within a year, Mythos will probably look quite dumb (relative to other new models). And others may release openly available or unguardrailed models of Mythos-level capabilities.
We started Project Glasswing because capabilities like Mythos Preview's won't stay rare, or stay in careful hands. We are bringing it to defenders as fast as we responsibly can, while working to figure out, for example, the right safeguards and patching & disclosure processes.
Also, to be clear, compute has never been a limiter in our rollout.
Expect a fuller update on our Glasswing work in the coming days.
XBOW report: https://xbow.com/blog/mythos-offensive-security-xbow-evaluation
UK AISI report: https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing
https://www.wsj.com/tech/ai/anthropic-mythos-apple-macos-bug-339da403
Mythos has cracked MacOS. It took five days.
One note of caution about the UK AISI cybersecurity benchmarks (and in particular TLO) is that neither Mythos nor GPT-5.5 was given the chance to "plateau" on the benchmark to measure the model's ceiling.
Mythos performs better on this benchmark per token, but GPT-5.5 outperforms Mythos per dollar spent.
Very important update from UK AISI. This is a meaningful change from the previous report. Here’s what the new data would look like for “Mythos Preview (new)” with $ on the x-axis:
From Calif:
@AndrewCurran_ Direct link to the blog from @calif_io https://blog.calif.io/p/first-public-kernel-memory-corruption
'Mythos Preview is powerful: once it has learned how to attack a class of problems, it generalizes to nearly any problem in that class.'
https://blog.calif.io/p/first-public-kernel-memory-corruption
https://blog.calif.io/p/first-public-kernel-memory-corruption
narrator voice: it was not gpt-2
Mythos has cracked MacOS. It took five days.
This means that the choice of x-axis is meaningful. Performance rankings can change based on whether you choose tokens, $, seconds, or something else on the x-axis. Each one, on its own, is incomplete. So it’s helpful to have the full picture.
@tymrtn Calif says they will publish a 55 page report after Apple ships a fix.
XBOW has a good article on what happens when you measure performance with: - a fixed “action” budget - a fixed token budget - a fixed $ budget
Uh oh. Captain, its just thursday
Mythos has cracked MacOS. It took five days.
Whaaaat. No way. That’s pretty cool if they’re talking about the Secure Enclave Processor, because then it means that Apple’s key invalidation per OS install is insufficient security. Old Intels used to write zeroes to make data “unrecoverable” between OS installs. That is cool it found a way through it, because it also means that Apple can better test future OS releases. I really hope they or Apple do a write up on it, once it’s patched. Because Apple has staked a lot on the robustness of the Secure Enclave.
The new (correct) Mythos Preview checkpoint outperforms other models on the TLO cyber range per token, though not per dollar.
UK AISI did not spend enough tokens to see performance plateau for either model. Both would have kept getting better if given more tokens.
@AndrewCurran_ Direct link to the blog from @calif_io https://blog.calif.io/p/first-public-kernel-memory-corruption
@AndrewCurran_ Now I understand why the Microsoft engineer decided to become a goose farmer
It’s a great read! But one caution when reading XBOW’s article: some charts have % on the y-axis, others have odds. Odds can make small differences close to 100% (or 0%) appear larger than a %-based chart. https://xbow.com/blog/mythos-offensive-security-xbow-evaluation
