The State of AI Cybersecurity 2026 report reveals a widening gap between the rapid adoption of autonomous AI agents and organisations’ ability to secure them. The research analyses 490 conference submissions from 501 speakers, offering insight into emerging cybersecurity priorities.
The report, prepared by Rob T. Lee of SANS Institute and Indre Altman of Crosspoint Capital Partners, finds that 71% of submissions focus on AI agents, compared with 66% in March 2026. Agentic systems have become the industry's dominant security concern.
Security research is increasingly focused on attacking AI systems rather than merely governing their deployment. Offensive research accounts for 46% of submissions, while defensive research has declined to 30%, exposing a potentially dangerous imbalance.
The most striking differentiation is between research interest and operational readiness. Although prompt injection, jailbreaking and vulnerability discovery attract considerable attention, only 6–8% of relevant submissions report results from production systems.
By contrast, Security Operations Centre and incident response research demonstrates comparatively greater operational maturity, with 21% reporting production results. This suggests that established security operations are adapting faster than specialised agent-defence mechanisms.
Key Highlights
REPORT AT A GLANCE
2026
490
Research submissions
71%
Focus on AI agents
12%
Agentic AI dominates: 348 of 490 submissions examine AI agents.
Offensive research leads: 46% of submissions are offense-led, versus 30% defense-led.
Production readiness remains limited: Just 12% of submissions report production-system results.
Identity becomes critical: Identity and authorisation appear in 34% of submissions.
Open-weight adoption accelerates: Mentions rise from 9% to 23%.
Governance loses attention: Governance-related submissions decline substantially.
Threat intelligence remains scarce: Only 8% model an attacker using AI.
Identity and authorisation have emerged as critical priorities, appearing in 34% of submissions. Autonomous agents require secure delegation, least-privilege permissions, continuous verification and accountability when accessing enterprise applications or sensitive information.
Openweight AI models are also gaining prominence, rising from 9% to 23% of submissions. Cost efficiency, data residency, privacy and deployment flexibility are driving adoption, while introducing additional supply-chain and model-governance considerations.
Another concern is limited visibility into real-world adversaries. Only 8% of submissions examine attackers using AI, suggesting that much of the discussion centres on theoretical attacks, demonstrations and simulated vulnerabilities rather than documented criminal operations.
Governance receives comparatively less attention, falling from approximately 32% to 13% of submissions. This shift may reflect greater emphasis on practical technical controls, but it also exposes the need to connect operational security with regulatory accountability.
The central lesson is that securing AI models is no longer sufficient. Enterprises must protect the entire agentic ecosystem—including identity, tools, memory, plugins, execution environments and data flows—through continuous monitoring, enforceable permissions and measurable production security.
