Mythos AI: Get your AppSec game on

Anthropic has provided a frothy couple of weeks for teams following the AI space. First came a leak of marketing materials for the company’s impending release of its most advanced model to date, Claude Mythos (code name: Capybara). That was followed almost immediately by the exposure of the source code for Claude Code. Much of the coverage has felt a bit like TMZ for tech — light on details and heavy on hot takes. 

Some conspiracy-minded tech folk even wondered whether it wasn’t all intentional, in pursuit of publicity for Mythos. Most security strategists, however, were more interested in the substance of what was released — and what it means for security road maps.

For application security (AppSec) veterans, Mythos looms as a big boost in the acceleration of vulnerability discovery — and exploit development. In confirming the leak, Anthropic told Fortune that Mythos will be a “step change” in reasoning and cybersecurity research capabilities. 

Anthropic’s leaked blog post said:

“Although Mythos is currently far ahead of any other AI models in cyber capabilities, it presages an upcoming wave of models that can exploit vulnerabilities in ways that outpace the efforts of defenders.” 

Anthropic has now launched Claude Mythos Preview to select partners as part of a responsible disclosure process, which includes Project Glasswing, "an effort to use Mythos Preview to help secure the world’s most critical software, and to prepare the industry for the practices we all will need to adopt to keep ahead of cyberattackers," the company said.

Here’s what you need to know about Claude Mythos — and what it portends for AppSec teams.

[ See webinar: Stop Trusting Packages — Start Verifying Them ]

Vulnpocalypse: What it is — and why it matters

The Anthropic leaks sparked concerns by researchers at the [un]prompted AI security conference recently. Heather Adkins, vice president of security engineering at Google, was among those alarmed about “something close to a cataclysmic increase” in vulnerability discovery and disclosure. John “Four” Flynn, vice president of security and privacy at Google Deepmind, said that what he dubbed the “vulnpocalypse” has already begun . 

Also at [un]prompted, Adam Laurie, CISO at Alpitronic, demonstrated how he used Claude to automate a hardware hacking lab — and own an LPC chip in seven minutes. Adam Křivka, AI security engineer at AISLE, showcased an AI system that discovered 12 zero-day vulnerabilities in the OpenSSL codebase. And Sergej Epp, CISO at Systdig, demoed an AI-assisted attack that moved from stolen credentials to full administrator access in a target AWS environment in just eight minutes.

Anthropic researchers were also on the bill, giving attendees a peek at what Mythos would bring. Nicolas Carlini, an Anthropic research scientist, said new state-of-the-art AI models are finding zero-days even in software projects that have been extensively tested for decades.

“LLMs can autonomously, and without fancy scaffolding, find and exploit zero-days in critical software. And they are getting good scarily fast. These new capabilities will alter the threat landscape and require [that] we rethink security in the coming years.’
—Nicolas Carlini

In short, seasoned security researchers and big thinkers say we are on the precipice of a huge shakeup in how vulnerabilities are found, exploited, and remediated.

AI is shifting into high gear

Security researcher and software developer Thomas Ptacek wrote in a think piece recently, Vulnerability Research Is Cooked:

“You can’t design a better problem for an LLM agent than exploitation research. Vulnerabilities are found by pattern-matching bug classes and constraint-solving for reachability and exploitability. Precisely the implicit search problems that LLMs are most gifted at solving. Agents are uncannily skilled at software development, and vulnerabilities are at the apex of that skill.”
—Thomas Ptacek

Phil Venables, a former Google CISO and now a partner at Ballistic Ventures, said that a year ago he had expected AI to impact cybersecurity only incrementally. Now he thinks the negative impacts will be bigger and more immediate. Nonetheless, he expects an even larger positive impact as defenses are improved by applying AI models and agentic capabilities to automated vulnerability remediation. 

“I am short-term pessimistic but wildly long-term optimistic.”
—Phil Venables

Others were also optimistic. Deepmind’s Flynn told the [un]prompted gathering that tools such as Google’s Code Mender could turn back the vulnpocalypse. Code Mender is an autonomous agent designed to debug and fix complex vulnerabilities, and it is just one example of automated remediation tools that defenders expect will be augmenting the AppSec tool stack soon.

Such defensive tools will help, wrote Marcus Hutchins, principal threat researcher for Expel. But even more effective may be the economics of finding and fixing bugs.

“Defenders are the ones with all the resources. They’re the ones building multi-billion dollar AI models specifically for auditing software, which criminals can’t even come close to finding the funding to build.”
—Marcus Hutchins

The agentic software factory is near

The Anthropic leaks also highlighted that the agentic AI attack surface is large and growing larger. AI security researcher Jiten Oswal wrote that the leaked code included multiple feature flags.

“The leak unveiled that Anthropic is sitting on a treasure trove of unreleased, fully-built agentic features.”
—Jiten Oswal

It all points to the next generation of large language models being purpose-built for agentic action, and Nipun Gupta, founder of the agentic AI security firm Optimus Labs, thinks the agentic software factory can’t be too far off.

“Which also means your agents become a new attack surface when they have so much capability, when they have so much to do with not just the setup and the builder collaboration, but also making and taking actions on your behalf. So your software supply chain is at much, much greater risk.”
—Nipun Gupta

Challenge your assumptions

The spread of agentic action in development pipelines makes many security road map assumptions nonviable, said Chris Hughes of ResilientCyber in a recent post. 

“The human-in-the-loop is not functioning as a meaningful safety control. It is a formality that users power through to maintain their workflow.”
—Chris Hughes

Advanced models and agent autonomy are going to open up whole new classes of risks, Gupta said, especially in organizations that have granted agents the same level of access to systems that an experienced red teamer might have.

“We used to compromise the machine and install a lot of these products that would allow us to maintain persistent access to the victim’s machine. Now we don’t need to because agents have already done that. So all I need to do is have a prompt injection, compromise the agent, and then I have persistent access forever.”
—Nipun Gupta

Observability and visibility matter more than ever

At the end of the day, what’s needed is just better security overall. When applied to agentic environments, the fundamental patterns still look familiar: establishing strong visibility, instituting controls that restrict either agentic privileges or actions, and building up layered security mechanisms that backstop one another.

Dhaval Shah, senior director of product management for ReversingLabs, said security leads should be thinking about this through the lens of zero trust and deep artifact assessment. 

“They need to accept that they cannot rely entirely on preventative scanning for AI agents. Because the inputs and outputs are natural language and highly dynamic, signature-based detection will fail.”
—Dhaval Shah

ResilientCyber’s Hughes said that no matter the agentic use case, organizations should be doubling down on visibility and observability to understand where agents exist within the infrastructure and what they’re authorized to do. From there, they should institute a mix of both deterministic and probabilistic controls.

“Build hard boundaries, layer deterministic and probabilistic controls, invest in runtime visibility and treat agent permissions as an infrastructure problem, not a user behavior problem. We need to build security programs for how humans actually interact with them, not how we wish they would.”
—Chris Hughes

Learn how RL's free Spectra Assure Community can help your development and AppSec teams get deep insights into your software supply chain via binary analysis.