A follow-up to my previous post on Mythos Preview. The AI Security Institute (AISI) has published a very interesting analysis of Mythos Preview. Very interesting because:
- AISI is "a mission-driven research organisation in the heart of the UK government". Its reports are clearly much more credible than claims of the form "our last product is too strong to give you, believe us" by a private US company, that is currently losing lot of money, that is fiercely battling against other companies in the AI arena, that is extremely good at fuelling hype about their products and capabilities.
- They consider complete cybersecurity tasks, i.e. CTF (capture the flag) competitions and attacks to a simulated organization.
- They compare the behavior of different models for a given "token budget".
Not surprisingly, Mythos Preview is indeed very good and better than previous models, but it is definitely not the coming Apocalipsis. In particular, it is the first tool that has completed a certain benchmark defined by the AISI: "a 32-step corporate network attack simulation spanning initial reconnaissance through to full network takeover, which we estimate to require humans 20 hours to complete". However, as pointed our by AISI itself (bold and bulleted list mine):
- Mythos Preview’s success on one cyber range indicates that is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained.
- However, our ranges have important differences from real-world environments that make them easier targets. They lack security features that are often present, such as active defenders and defensive tooling.
- There are also no penalties for the model for undertaking actions that would trigger security alerts.
- This means we cannot say for sure whether Mythos Preview would be able to attack well-defended systems.
As Gary Marcus put it, "One hopes that by now no mission-critical infrastructure is “small, weakly defended, and vulnerable” with ready network access."
AI-based tools are quickly improving their capabilities and while they have not yet influenced the behavior of real world attackers (topic for a future post), they will certainly have important implications for cybersecurity in the near term (idem). However, we should all stop propagating any claim by AI companies, acting as unwitting and unpaid employees of their marketing teams, without first giving it critical scrutiny.
The Guardian has a very interesting analysis of the recent announcement of Mythos Preview.
Commenti