AI is changing the speed, scale, and economics of attacks. What worked for security in a world of human actors struggles in a world of AI, agents, autonomy, and continuous adaptation. We believe security needs a new Cyber Stack. One that will empower defenders with real-time protections, defending against AI with AI. That conviction led us to build Project Perception, a new agentic security system designed for the realities of AI, in public preview August 3. What sets Perception apart is its combination of Security Context, a Multi-Model Harness, and Specialized Security Agents. Earlier this year, we introduced MDASH, our software vulnerability multi-modal team of agents. MDASH launched beating Mythos, GPT and Gemini on CyberGym, the gold standard benchmark for evaluating how systems reason over large codebases to find real vulnerabilities. MDASH reinforces our belief in a multi-modal approach and that you deliver the best results by applying the right model, to the right task, at the right time. And this is why we are partnering with Mustafa Suleyman and the MAI team to continue to push the limits with MAI Cyber-1-Flash inside MDASH. Together, MDASH with MAI-Cyber-1-Flash deliver 96% on CyberGym (+12 pt above Mythos) at 50% of the cost when compared to our existing MDASH offering (MDASH + GPT 5.4 + 5.4 mini + 5.3 Codex). There should not be a choice between innovation and safety. Security should be an accelerator, not a blocker. Welcome to Perception. https://lnkd.in/gX7XN4xZ
AI is fundamentally reshaping the threat landscape, and traditional security models simply can’t keep up with autonomous, adaptive attacks. The shift toward agentic security systems like Project Perception — combining deep security context, multi‑model orchestration, and specialized agents — is exactly what modern defense requires. MDASH’s performance on CyberGym shows how powerful the right model applied at the right moment can be. Security must evolve at the same pace as AI, and this direction makes that possible. Great insights.
This is a strong direction. The combination of security context, specialized agents, and a multi-model harness recognizes that no single model is likely to remain best across every threat, codebase, and stage of investigation. I’m curious how model selection and agent authority are governed when speed and confidence point in different directions—for example, when the fastest model produces a high-confidence recommendation that conflicts with a slower specialist. One additional requirement is cross-model evidence integrity. In a multi-model defense system, each handoff should preserve the source evidence, assumptions, confidence level, and decision history—not just the conclusion. Otherwise, orchestration can improve performance while making accountability and post-incident reconstruction more difficult.
I wonder if the next frontier isn’t larger models, but better AI operating systems, systems that know which model to use, when to use it, how to verify the output, and when to ask for human input. The orchestration layer may end up being as important as the models themselves.
Kapil Panwar
1w
Hayete, the multi-model harness approach is exactly what production AI needs. At Gyannetra, we have found that no single model is optimal for every task — routing the right query to the right model at the right time is what separates reliable systems from fragile ones. The 50% cost reduction at higher accuracy is a compelling data point. Would love to hear more about how Perception handles model routing decisions at scale.
Hayete Gallot, a challenge that may define the next stage of AI-powered security is not only making agents detect more threats, but enabling them to understand when a signal requires action and when restraint is the safer decision. Stronger feedback loops with security experts could help agents develop this judgment over time, improving reliability while maintaining trust in complex environments.
Yandong Xiao
2w
The security context layer may be one of the most important parts of this architecture. Specialized agents can move quickly, but dependable production use still requires clear boundaries, traceable decisions, and reliable controls.
Shan Huang
2w
The 50% cost drop may matter more in production than the benchmark win. Security agents only change the operating model when teams can afford to run them continuously across noisy environments. Where is the multi-model harness saving the most today: triage, code reasoning, or validation?
An exciting direction. AI is undoubtedly changing both the attack and defence landscape, and defenders will increasingly need AI operating at machine speed to keep pace. One additional thought: while AI-powered security agents will strengthen detection and response, sustainable cyber resilience will still depend on the broader enterprise operating model. Governance, identity, trusted data, human accountability and well-defined decision rights remain essential. The most effective organisations will combine AI-native security capabilities with equally mature governance and operational discipline.
Huge step forward. I’m spending a lot of time thinking about what “security at AI speed” really means — and this moves the conversation forward in a very practical way: not just AI agents, but the data foundation needed to ground them. The harder part may be governance. Companies will need to evolve governance to operate at AI speed too, because technology and governance have to move together. Attackers don’t care how we’re organized, where the silos are, or who owns which process. I’ve been writing quite a bit about this lately, and honestly, this is probably the most interesting ride I’ve seen in security in a long time.