Anthropic is keeping its unreleased Mythos AI model private after it demonstrated the ability to automate multi‑step cyber‑attacks.
If the system can discover and exploit software flaws faster than patches can be applied, it could reshape global cyber‑defence dynamics and raise the stakes for governments and businesses.
Mythos is designed to automate exploit chaining and conduct simulated multi‑step attacks on software worldwide. Independent testing reported that the model succeeded in expert‑level cybersecurity challenges roughly 73% of the time, showing a high level of competence in complex attack scenarios[1].
Anthropic said the system has already uncovered thousands of critical software vulnerabilities, a claim that underscores the model’s potential to surface flaws that human researchers might miss[2]. The company said that keeping the model private allows it to study these findings without exposing the techniques to malicious actors.
Security experts are split on the level of risk. Some analysts said the model’s capabilities matter and could accelerate the discovery of exploit chains, while others said the most alarming risk assessments may be overstated[4]. The 73% success rate cited by independent testing has fueled the debate, with some seeing it as evidence of a high‑risk tool.
Anthropic said the decision to withhold Mythos is based on the model’s perceived danger to the public, saying that releasing it could enable automated attacks that outpace defensive measures[3]. The company has not disclosed a timeline for a broader release, opting instead for internal use and controlled testing.
Industry leaders, including major cloud providers, have expressed caution, noting that the emergence of such AI‑driven attack tools could force a reevaluation of current security practices. As AI continues to advance, the balance between innovation and safety remains a central concern.
**What this means**: Mythos illustrates how AI can both accelerate vulnerability discovery and create new avenues for large‑scale cyber‑attacks. Policymakers and security teams will need to develop frameworks that allow beneficial AI research while mitigating the risk of powerful tools falling into hostile hands.
“Mythos succeeded in expert‑level cybersecurity challenges roughly 73% of the time.”
Mythos illustrates how AI can both accelerate vulnerability discovery and create new avenues for large‑scale cyber‑attacks. Policymakers and security teams will need to develop frameworks that allow beneficial AI research while mitigating the risk of powerful tools falling into hostile hands.
