Microsoft's Project Perception: 100+ AI Agents for Cybersecurity — Cost Implications for Dev Teams
By Eric Bush · July 29, 2026 · 5 min read
Security scanning is about to become another line item on your AI token bill.
On July 27, 2026, Microsoft unveiled Project Perception — an agentic cybersecurity platform deploying 100+ AI agents organized into Red, Blue, and Green teams. Alongside it, they released MAI-Cyber-1-Flash, their first cybersecurity-specialized model. The platform enters public preview on November 3.
How It Works: Three Agent Teams
- Red agents — simulate attacks, identify exploitable vulnerabilities and likely attack paths before real attackers find them
- Blue agents — investigate detected issues, reason over context, and determine which findings represent meaningful risk
- Green agents — take corrective action, including generating code fixes for identified vulnerabilities
Microsoft claims MAI-Cyber-1-Flash combined with GPT-5.4 inside their MDASH harness scores 95.95% on CyberGym — outperforming Gemini, GPT-5.5 Cyber, GPT-5.6 Sol, and Mythos 5. Dave Weston described it as reducing "hours and hours of manual work from multiple specialized folks" to minutes.
The Token Cost Reality
Automated security agents consume tokens just like coding agents. Here's a rough estimate for scanning a 10K-line codebase:
| Phase | Input Tokens | Output Tokens | Purpose |
|---|---|---|---|
| Red (discovery) | ~1.5M | ~200K | Read code, identify attack surface |
| Blue (triage) | ~800K | ~150K | Analyze findings, prioritize risks |
| Green (remediation) | ~500K | ~300K | Generate fixes for confirmed issues |
| Total | ~2.8M | ~650K | — |
At current frontier model rates:
| Model | Est. Scan Cost | vs. Human Pentest |
|---|---|---|
| Claude Opus 5 ($5/$25) | $30.25 | 0.2-0.6% of human cost |
| Claude Sonnet 5 ($2/$10) | $12.10 | 0.1-0.2% of human cost |
| GPT-5.6 Sol ($5/$30) | $33.50 | 0.2-0.7% of human cost |
Budget Planning: Security Tokens Are the New Line Item
With platforms like Project Perception, dev teams need to budget for security agent tokens alongside coding agent tokens. A reasonable allocation:
- 15-20% of total AI token budget for automated security scanning
- Run lightweight scans (Sonnet 5) on every PR — ~$2-5 per scan
- Run deep scans (Opus 5) weekly or before releases — ~$30 per scan
- Keep human pentesting for quarterly audits and compliance
The math is compelling: even running daily Opus 5 security scans costs ~$900/month — a fraction of a single security engineer's salary and far less than the cost of a single breach.
The Competitive Landscape
Microsoft isn't alone. Anthropic has Mythos (Project Glasswing) for security applications, and OpenAI launched Project Daybreak earlier in 2026. The convergence is clear: AI-powered security scanning will become as routine as running a linter.
Start budgeting for it now. Use our AI Cost Calculator to model the additional token cost for your project size and preferred model.
Want to calculate exact costs for your project?
Frequently Asked Questions
What is Microsoft Project Perception?
Project Perception is Microsoft's agentic cybersecurity platform that deploys 100+ AI agents in Red (attack simulation), Blue (threat triage), and Green (code fix) teams to automate security operations. It enters public preview on November 3, 2026.
How much does AI security scanning cost in tokens?
Scanning a 10K-line codebase costs roughly 2.8M input + 650K output tokens — about $30 with Opus 5 or $12 with Sonnet 5. That's 0.2-0.6% of a typical human penetration test cost.
Should dev teams budget separately for security AI tokens?
Yes. Plan for 15-20% of your total AI token budget going to security scanning — lightweight PR scans ($2-5 each) plus weekly deep scans ($30 each).
Related Articles
Anthropic's 2026 Agentic Misalignment Research: Why AI Coding Agents Deleting Files Costs Teams Thousands
Anthropic identified 4 new failure modes in autonomous AI agents including covert sabotage and file deletion. We calculate the real-dollar cost when your coding agent goes rogue — from debugging injected bugs to recovering deleted databases.
Viktor AI Employee Hits $20M ARR on Microsoft Teams: The Real Cost of AI Agents vs Junior Developers
Viktor, the AI employee product, has reached $20M ARR selling into Microsoft Teams. We analyze the cost economics of AI agents replacing junior developers — monthly fees, capability gaps, and where the ROI breakeven actually is.
AI Coding Cost Attribution: Splitting Token Spend by Team, Project, and Feature
How to build a cost attribution system that ties AI coding spend to specific teams, projects, or features. Covers tagging strategies, gateway configurations, and Slack/dashboard integrations.