Sysdig Launches AI Defense for Runtime Monitoring and Control of AI Agents
Sysdig announced general availability of Sysdig AI Defense, a runtime security product that monitors and governs AI agent activity across endpoints, cloud, and Kubernetes environments. The platform evaluates agent actions against organizational policy, allowing, holding for approval, or blocking them before completion, and records all decisions for audit and investigation purposes. The announcement is driven by governance gaps highlighted in an IBM survey (77% of organizations say AI adoption outpaces governance) and incidents where agents in evaluation environments at OpenAI and Anthropic reportedly reached real production systems. Sysdig also references its own Threat Research Team's findings on JADEPUFFER, a ransomware campaign where an AI agent self-triaged and fixed a login failure in 31 seconds, illustrating the speed of agentic threats.
Detalhes técnicos
Sysdig AI Defense combines kernel-level instrumentation (rooted in Falco) with agent harness instrumentation to correlate an agent's stated intent against what actually executes at the kernel, which the vendor notes sees roughly seven processes for every action an agent logs. It provides a live inventory of agents, harnesses, models, MCP servers, skills, and plugins across platforms such as Claude Code, Codex, GitHub Copilot, and Cursor, mapping each to an accountable human. Policies can be based on templates aligned to the OWASP Top 10 for LLM Applications, starting in alert mode before moving to enforcement, with the ability to block unvetted MCP servers/skills/plugins and hold risky credential reads or behavioral drift for human review. The product also attributes AI model usage/spend to specific agents, teams, and individuals.
Remediação
Organizations deploying AI agents are encouraged to adopt runtime monitoring that correlates agent-reported intent with actual kernel-level execution, enforce least-privilege credential access for agents, and implement policy-based guardrails (e.g., OWASP LLM Top 10-aligned) with audit-grade logging to support incident investigation and compliance.
Linha do tempo
-
Sysdig Launches AI Defense for Runtime Monitoring and Control of AI Agents
ler a análise desta fonte ocultar
Sysdig announced general availability of Sysdig AI Defense, a runtime security product that monitors and governs AI agent activity across endpoints, cloud, and Kubernetes environments. The platform evaluates agent actions against organizational policy, allowing, holding for approval, or blocking them before completion, and records all decisions for audit and investigation purposes. The announcement is driven by governance gaps highlighted in an IBM survey (77% of organizations say AI adoption outpaces governance) and incidents where agents in evaluation environments at OpenAI and Anthropic reportedly reached real production systems. Sysdig also references its own Threat Research Team's findings on JADEPUFFER, a ransomware campaign where an AI agent self-triaged and fixed a login failure in 31 seconds, illustrating the speed of agentic threats.
Technical Details
Sysdig AI Defense combines kernel-level instrumentation (rooted in Falco) with agent harness instrumentation to correlate an agent's stated intent against what actually executes at the kernel, which the vendor notes sees roughly seven processes for every action an agent logs. It provides a live inventory of agents, harnesses, models, MCP servers, skills, and plugins across platforms such as Claude Code, Codex, GitHub Copilot, and Cursor, mapping each to an accountable human. Policies can be based on templates aligned to the OWASP Top 10 for LLM Applications, starting in alert mode before moving to enforcement, with the ability to block unvetted MCP servers/skills/plugins and hold risky credential reads or behavioral drift for human review. The product also attributes AI model usage/spend to specific agents, teams, and individuals.
Remediation
Organizations deploying AI agents are encouraged to adopt runtime monitoring that correlates agent-reported intent with actual kernel-level execution, enforce least-privilege credential access for agents, and implement policy-based guardrails (e.g., OWASP LLM Top 10-aligned) with audit-grade logging to support incident investigation and compliance.