AI has got better at hacking—how big a risk is it?

Anthropic's Decision to Withhold Mythos: Anthropic has decided not to release its latest AI model, "Mythos," to the public due to its exceptional capability in autonomously discovering software vulnerabilities and exploiting them to crash systems or gain unauthorized access.

• The model represents a shift from a coding assistant requiring human oversight to an "automated hacker" capable of identifying and executing exploits with minimal human intervention.
• Unlike standard consumer models, the Mythos preview lacks the safety filters and refusal mechanisms typically trained into AI to prevent the generation of malicious code.

Escalation in Vulnerability Lifecycle Speed: Data indicates a rapid decrease in the time window between the discovery of a software vulnerability and its exploitation in active attacks.

• In 2018, the average delay between disclosure and exploitation was 2.3 years.
• Current delays have plummeted to approximately 20 hours.
• At the current rate of acceleration, this window is projected to shrink to one minute by 2028.

Demonstrated Capability via OpenBSD Exploit: Mythos successfully identified and generated an exploit for a vulnerability in the OpenBSD operating system that had remained undiscovered for 27 years.

• The vulnerability could be triggered by a single message, causing a "kernel panic" (a total system freeze).
• The flaw was resolved with a one-line code fix, but the difficulty in identifying it previously highlights the model's advanced pattern recognition.
• Anthropic utilized "responsible disclosure," sharing the details only after the vulnerability was patched.

Root Cause of Capabilities: Mythos's efficacy is attributed to the scaling laws of AI rather than specialized cybersecurity training data.

• The model is larger than its predecessor, Opus, with increased compute resources ("flops") and data, making it superior at general tasks including security auditing.
• It was not specifically trained on a dataset of existing cyber threats to identify similar bugs, but rather applied general intelligence to find weaknesses.

Broader Security Implications: The emergence of such models raises concerns regarding the future balance between attackers and defenders.

• There is a risk that open-source models will soon achieve similar "insane hacking skills," potentially democratizing high-level cyberattacks.
• The distinction between "hacking" (subverting intent) and "malicious hacking" is acknowledged, though the discussion focuses on the colloquial use of the term regarding unauthorized system access.