EXPLORINEER
en

KR — 전자신문

[AI Guarding AI] Microsoft Extends Project Perception to Protect Internal AI Agents, Not Just External Threats

Security1 min read

Microsoft has described a new AI security system called Project Perception. It is built to watch and protect not only outside AI threats but also AI agents working inside a company.

What happened

Microsoft introduced Project Perception in July as an 'agentic security' system. It combines signals, security context, AI models, and special AI agents into one learning defense system. The goal is to spot risks, decide what they mean, and take protective steps across a company's digital systems. Microsoft now says this protection must also cover AI agents that work inside a company.

Key facts

  • Project Perception has three special agents: a Red Team agent, a Blue Team agent, and a Green Team agent.
  • The Red Team agent finds possible break-in paths before attackers can use them.
  • The Blue Team agent studies security context to find real risks.
  • The Green Team agent makes fixes to strengthen defenses across the environment.
  • A Guardian agent watches other agents' actions when risk rises and responds in real time.

Background

Old AI security mainly focused on outside threats, where attackers use AI to launch attacks. Now, AI agents can enter a company's data, tools, and systems to do real work and help make decisions. Microsoft calls these company-run agents 'agentic insiders.' Even agents a company builds itself need constant checks on what they do and how risky they are.

Why it matters

AI agents can act quickly, so security teams need tools that work at the same speed. Microsoft says people must keep final control over important decisions. This approach aims to help security staff focus on the most serious risks, not just handle more alerts. It also gives logs that show what caused a problem, so teams can adjust their policies and tools.

What to watch

  • How well Project Perception works for companies in real use.
  • Whether Microsoft's layered protection covers identity, data, cloud, devices, networks, and runtime well.
  • How companies balance fast automatic responses with human control.

Glossary

Project Perception
Microsoft's agentic security system that uses several AI agents to find and respond to risks.
Red Team agent
An AI agent that looks for possible attack paths before criminals can use them.
Blue Team agent
An AI agent that studies security context to decide which risks are real.
Green Team agent
An AI agent that carries out fixes to make the whole system safer.
Guardian agent
An AI agent that watches other agents' actions and responds in real time when risk is high.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

More in this topic

SecurityITmedia

GMO Confirms Attacker Published Information; Investigating Whether Customer Data Is Included, Considering Additional Condolence Payments

GMO Internet Group said that at 12:30 p.m. on October 27 it confirmed the publication, apparently by an attacker, of information that appears to concern customers, related to unauthorized access to its FX auto-trading VPS "Onamae.com Desktop Cloud." The company is still investigating the contents, which may include inaccurate information, and is considering additional condolence payments for affected users whose information was published, beyond the flat 5,000-yen payment announced earlier.

SecurityITmedia

Times Car Asks How the ¥3,000 Apology Payment Was Calculated; Park24 Responds

Park24, the operator of the car-sharing service Times Car, told ITmedia NEWS on October 27 that the ¥3,000-equivalent apology payment to affected users was set by comprehensively weighing the response options available to users and the burdens involved, though it declined to disclose the specific calculation basis. The company said the payment will be made as cash or vouchers, but the form and payment timing are still under consideration, and details will be announced once decided. The payment covers 1,464,349... (see note) affected records, and the company expects to book a ¥5.5 billion special loss for the payment and related response costs.

Security전자신문

[AI Against AI] Piolink: From AI-Driven Vulnerability Verification to Blocking Hacks

As AI-driven hacking automates vulnerability discovery, attack-path design, and intrusion attempts, enterprise security teams face growing pressure. Piolink proposes a multi-layered defense: its agentic AI-based CTEM platform "BlueHunter" verifies real exploitability to set priorities, its web firewall "WebFront-K" blocks automated attacks in real time, and 24-hour security monitoring tracks attack continuity. The company said BlueHunter found 28 undisclosed vulnerabilities in 10 Linux open-source services and one CMS, with two assigned CVE numbers and disclosure planned for December.

SecurityITmedia

GMO to Provide New PCs to Customers Hit by Ransomware; Criminal Group Reportedly Publishes One Customer's Data

GMO Internet Group announced it will provide replacement PCs with equivalent performance to customers whose PCs were hit by ransomware through unauthorized access to "お名前.com デスクトップクラウド." The company said that as of 11:59 a.m. on October 27, no ransomware or other malware had been confirmed on devices already collected and analyzed, while the criminal group published what it claimed was one customer's personal information, which GMO said may contain inaccurate content and urged people not to access or spread it.