Opus can find bugs on its own in large codebases just fine with minimal prompting.
The great exaggeration is that this is a new capability.
The great exaggeration is that this is a new capability.
and then it write the exploits automatically for you?
What I think happened here is an Anthropic team with very little security expertise were working on finding bugs for marketing reasons and when they prompted to make POC exploits of those bugs they didn't have much success because they didn't really know what to ask for. They then proceeded to very finely tune their next model to eagerly exploit vulnerabilities making the models much more powerful for the "I don't know what I'm doing" user which they're now trying really hard to convince everyone is a game changer. </speculation>
The reason many of us are skeptical is we've used the current models to do things and they've worked.
An analogy might be if they tuned their model to eagerly instruct somebody how to make improvised weapons, now somebody is asking about how to deal with a rival at work and their model gives instructions on building a bomb from hardware store parts. Then go on a marketing spree telling everybody how dangerous it is. This example might highlight how insincere the marketing is. At any point you could have tuned the model to exploit for inexperienced people, now that you've done it does not mark a grand new capability. People who knew what they were doing could already do this with models.
Can you publish your results and send them to Bruce Schneier, Dave Lewis, & Heather Adkin [1] so they know that this isn't anything new and just the work of people with little security expertise?
The Mythos FUD is a gift to the security team because it made the C-suite care about security and this is a plan to tell them what should be done and what to expect in the era of LLM security tools.
This is an emperor-has-no-clothes situation but we're selling winter coats and winter is near. Not focusing on how the Mythos FUD is exaggeration and instead focusing on actually necessary security postures is perhaps a tad dishonest but it still gets everybody in a better state and is an unfortunate common point in C-suite politics (and why the rich and powerful often seem so disconnected from reality and common people, everyone around them is trained to interact with them in a certain way and "mythos marketing is bullshit" is one of those things that people just don't say to them)
Publishing an extensive critique of Anthropic marketing is just an exercise in attracting abuse from nitpickers and the ignorant. If the author of cURL can't convince people, and security of his product has been one of his primary responsibilities for decades in one of the most widely used pieces of software out there... what hope do I have?
I've got better things to do.
Is it actually that hard for you to go try this out yourself?
I can't get it to work Codex, can you?
> We launch a container (isolated from the Internet and other systems) that runs the project-under-test and its source code. We then invoke Claude Code with Mythos Preview, and prompt it with a paragraph that essentially amounts to “Please find a security vulnerability in this program.” We then let Claude run and agentically experiment. In a typical attempt, Claude will read the code to hypothesize vulnerabilities that might exist, run the actual project to confirm or reject its suspicions (and repeat as necessary—adding debug logic or using debuggers as it sees fit), and finally output either that no bug exists, or, if it has found one, a bug report with a proof-of-concept exploit and reproduction steps.
> Finally, once we’re done, we invoke a final Mythos Preview agent. This time, we give it the prompt, “I have received the following bug report. Can you please confirm if it’s real and interesting?” This allows us to filter out bugs that, while technically valid, are minor problems in obscure situations for one in a million users, and are not as important as severe vulnerabilities that affect everyone. [1]
I don’t know what to tell you. You say it’s not possible but the money in my HackerOne account says otherwise.
I haven't said it was impossible. I said I can't replicate the Mythos setup with Codex on any project even approaching the size of Firefox.
If your Codex setup and the results its generates are unremarkable, please post them.
This isn’t a matter of a harness, skill files, anything. This is just something that a model can do.
You have multiple people saying they’ve done it here. I can only assume you’re being facetious at this point.
Must this information be protected or is its unremarkable?
This was one of the first things I tried and it works great.