Following their disclosure of the first AI-orchestrated cyber espionage campaign, Anthropic isn't just sounding the alarm—they're building the solution. Their latest initiative focuses on leveraging AI to strengthen defensive cybersecurity capabilities, marking what they call an "inflection point" in the cyber ecosystem.
The Defensive Advantage
Anthropic's research team has deliberately enhanced Claude Sonnet 4.5's ability to perform critical defensive tasks: discovering vulnerabilities in code, patching security flaws, and testing system weaknesses. The results are impressive. Claude Sonnet 4.5 now matches or exceeds the performance of their more expensive Opus 4.1 model in cybersecurity tasks, achieving a 76.5% success rate on complex CTF challenges when given multiple attempts—double the success rate from just six months ago.
In real-world testing using the CyberGym benchmark, Sonnet 4.5 successfully identified known vulnerabilities in 66.7% of tested programs and discovered new, previously unknown vulnerabilities in 33% of open-source projects when given 30 attempts per task. These aren't theoretical exercises—some of Claude's generated patches have been merged into actual open-source repositories.
Key Takeaways for Security Teams
Accessibility matters: At roughly $45 to run 30 vulnerability discovery attempts on a single target, AI-assisted security testing is becoming economically viable for organizations of all sizes.
Speed and scale: Tasks that would take skilled security professionals hours can now be completed in minutes. One complex challenge involving network traffic analysis, malware extraction, and decryption took Claude 38 minutes—work estimated to require at least an hour for a human expert.
Integration opportunities: Early adopters are already seeing results. HackerOne reported a 44% reduction in vulnerability intake time with 25% improved accuracy. The technology is ready for experimentation in SOC automation, SIEM analysis, and secure development pipelines.
The Urgency
Anthropic's message is clear: we cannot cede the AI advantage to attackers. With AI models now demonstrating practical utility for both offensive and defensive cyber operations, security teams need to start experimenting with these tools. The shift from theoretical capability to operational reality has already happened.
The question isn't whether AI will transform cybersecurity—it's whether defenders will adopt it fast enough to maintain the advantage.
Read the full research article: Building AI for cyber defenders