Microsoft extends Project Perception with Dynamic Permissions, risk scoring and Guardian Agents to contain rogue enterprise AI agents
Matéria consolidada. Reúne 2 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 2 fontes trouxeram em 24/10/2026 00:09; os links antigos trazem para cá.
Fontes: darkreading.com ↗ securityweek.com ↗
Microsoft announced new controls to monitor and restrict potentially rogue AI agents in enterprise environments, arguing that organizations should treat agents as insiders rather than trusted software (securityweek.com). The capabilities extend Project Perception, the agentic security system launched in July, which was initially pitched as a defense against external AI-driven attacks (securityweek.com). Hayete Gallot warned that agents are frequently overprivileged, with access to data, tools and systems, and can act in critical processes. Nadim Abdo clarified that the insider label does not mean agents are malicious, but that they can make mistakes, be manipulated or be compromised (securityweek.com). On 2026-10-23, Dark Reading described the addition of Guardian Agents to monitor and contain anomalous agents, and quoted experts who warn that policing unpredictable agents with other agents may multiply risk rather than contain it (darkreading.com). This is product news with no direct threat or incident attached (securityweek.com). The Dark Reading text available is only a short summary, so it gives no technical specifics, availability details or names of the experts.
Detalhes técnicos
- Dynamic Permissions: agents start with no or minimal permissions. Access is granted per task and expires or shrinks as the task changes or ends, and high-impact actions require human review (securityweek.com).
- Dynamic Agent Risk: a continuously updated per-agent risk score built from agent activity plus identity, data, endpoint, cloud, network and runtime signals. It factors in security configuration, reachable resources, potential damage and deviation from intended behavior (securityweek.com).
- Dynamic Conditional Access: brings Entra Conditional Access to agents. It can reduce permissions, allow read-only access, limit data and network reach, require human approval, or isolate the runtime or halt execution at high risk (securityweek.com).
- Playbooks and Guardian Agents: humans set rules and approval limits, Playbooks define the desired security response, and Guardian Agents investigate risk shifts and suggest or enforce containment (securityweek.com). Dark Reading frames Guardian Agents as monitoring and containing anomalous agents as a form of insider threat (darkreading.com).
- Agent discovery expands by 35 local agent types, for a total of 60, including Claude Code, OpenAI Codex and GitHub Copilot (securityweek.com).
Remediação
SecurityWeek frames the guidance as least-privilege, per-task access for agents, continuous risk scoring, human approval for consequential actions, and containment of agents whose risk rises (securityweek.com). Dark Reading's expert caveat suggests that defenders should not rely on automated Guardian Agents alone, and should keep human oversight of the controls themselves (darkreading.com).
Linha do tempo
-
Microsoft's Project Perception adds Guardian Agents to monitor rogue AI agents; experts warn of compounding risk
ler a análise desta fonte ocultar
Microsoft is extending its unified security stack with Project Perception, which adds Guardian Agents to monitor and contain anomalous AI agents operating inside enterprise environments. The approach treats misbehaving AI agents as a form of insider threat. Experts quoted by Dark Reading warn that policing unpredictable agents with other agents may multiply risk rather than contain it. The available article text is a short summary, so technical specifics, availability and the named experts are not stated.
-
Microsoft extends Project Perception with Dynamic Permissions and risk scoring to control rogue enterprise AI agents
ler a análise desta fonte ocultar
Microsoft announced new controls to monitor and restrict potentially rogue AI agents in enterprise environments, arguing that organizations should treat agents as insiders rather than trusted software. The capabilities extend Project Perception, the agentic security system launched in July, which was initially pitched as a defense against external AI-driven attacks. Hayete Gallot warned that agents are frequently overprivileged, with access to data, tools and systems, and can act in critical processes. Nadim Abdo clarified that the insider label does not mean agents are malicious, but that they can make mistakes, be manipulated or be compromised. This is product news with no direct threat or incident attached.
Technical Details
- Dynamic Permissions: agents start with no or minimal permissions. Access is granted per task and expires or shrinks as the task changes or ends, and high-impact actions require human review.
- Dynamic Agent Risk: a continuously updated per-agent risk score built from agent activity plus identity, data, endpoint, cloud, network and runtime signals. It factors in security configuration, reachable resources, potential damage and deviation from intended behavior.
- Dynamic Conditional Access: brings Entra Conditional Access to agents. It can reduce permissions, allow read-only access, limit data and network reach, require human approval, or isolate the runtime or halt execution at high risk.
- Playbooks and Guardian Agents: humans set rules and approval limits, Playbooks define the desired security response, and Guardian Agents investigate risk shifts and suggest or enforce containment.
- Agent discovery expands by 35 local agent types, for a total of 60, including Claude Code, OpenAI Codex and GitHub Copilot.
Remediation
The article frames the guidance as least-privilege, per-task access for agents, continuous risk scoring, human approval for consequential actions, and containment of agents whose risk rises.