Autonomous agents now drive threat research, intelligence gathering, and attack generation, dropping security scores 12.5% across all major models.
What happens when the attackers aren’t human? Today, CalypsoAI unveiled that autonomous AI agents – not human hackers – are now driving the future of cyber threats, running full-cycle threat research, intelligence gathering, and attack generation. The impact is already measurable: in the August CalypsoAI Security Index (CASI) Leaderboard, every leading AI model suffered a 12.5% drop in its security score, proving that agentic systems can compromise even the most “hardened” AI models.
These results are powered by Signature Attack Packs, a core capability of CalypsoAI’s Inference Red-Team product. Each month, Red-Team’s fully autonomous agent researches, generates, tests, and curates a new set of high-severity adversarial prompts. This process combines continuous threat intelligence with dynamic attack generation to expose real-world vulnerabilities at a speed and scale no human team can match.
It is often said that the best defense is a good offense. Our amazing team embraced the challenge of building a team of agents instead of linearly scaling the team. We now have the capability to conduct threat research, build intel and generate new and novel attacks which our customers benefit from in our Red-team and Defend products, all with agents doing the work.
The enhanced Inference Red-Team solution now incorporates Agentic Fingerprints, giving customers unprecedented observability into how attack agents behave – capturing decision-making, reasoning paths, and successful exploits for every campaign.
Meanwhile, Inference Defend gains Outcome Analysis, a new feature that provides clear visibility into why prompts and responses are flagged or blocked, eliminating guesswork and accelerating security response times.
“As enterprises rapidly adopt AI applications and agents in key functions, it is increasingly important to know exactly what these systems are doing at every step,” said Donnchadh Casey, CEO of CalypsoAI.”From our customers, we know that understanding where AI systems encounter vulnerabilities, make bad decisions, and why and how they fail is absolutely critical.”
To meet the needs of highly regulated industries and air-gapped environments, CalypsoAI has also introduced early access for Red-Team On-Premises, allowing enterprises to bring agentic red-teaming capabilities fully in-house without sacrificing speed or coverage.
CalypsoAI unveiled the upgraded Defend and Red-Team solutions at Black Hat USA 2025 in Las Vegas, the premier event for cybersecurity industry professionals, with over 22,000 attendees.
Explore AITechPark for the latest advancements in AI, IOT, Cybersecurity, AITech News, and insightful updates from industry experts!