CYBER DEFENSE AI
Copilot gets its own defense AI
Security Copilot, the generative-AI security assistant bundled with Microsoft 365 E5, has long leaned on third-party models. Microsoft's new in-house defense AI models arrive right as regulators circle its Copilot pricing.
A break from third-party dependence
Answering the question of who guards the guards
Security Copilot, the generative-AI security operations tool bundled into Microsoft 365 E5, has until now offloaded part of its threat detection and response processing to third-party models. Microsoft has announced a set of in-house cyber-defense AI models built to counter AI-driven attacks. The announcement itself functions as Microsoft's own answer to a question security teams have been asking for a while: which vendor should sit at the core of your defenses? As attackers increasingly weaponize generative AI, the discomfort of outsourcing the defensive model itself has been growing across the industry, not just around Security Copilot.
The timing is also hard to ignore. In the same stretch of late July 2026, Microsoft Copilot is under active antitrust scrutiny in multiple countries over a subscription price hike. According to reporting from The Register, the UK's Competition and Markets Authority (CMA), Italy's AGCM, and Australia's ACCC are all examining Copilot's pricing changes as a competition concern. That the in-house defense AI news landed in the middle of this headwind suggests it may double as a message to regulators and customers alike: the price increase comes with distinctive value. Read narrowly, this is a product announcement; read alongside the pricing probes, it looks more like a counter-message to the backlash.
What the numbers say about the field
Around the same time, OpenAI released Codex Security CLI, an open-source tool for vulnerability discovery, under an Apache-2.0 license — it has already picked up more than 1,500 stars on GitHub. This looks like the opening stage of a race among major AI vendors to stake out "AI for cyber defense and offense-testing" as their own turf. Microsoft's self-contained approach and OpenAI's open one represent genuinely different philosophies about how defensive AI should be built.
Engineers
Freed from third-party model dependence, detection accuracy may become less prone to unpredictable behavior shifts and API-cost swings whenever an upstream model changes. But an in-house, undisclosed model can also become a harder black box to debug when false positives spike.
Business leaders
Consolidating security operations under one vendor can simplify contracts and negotiation. But with antitrust probes already underway, deepening reliance on Microsoft is itself something a procurement review may now ask hard questions about.
Product/PMO
Since Security Copilot is folded into E5, defensive capability may improve at no extra cost. On the roadmap side, though, teams are left waiting to learn exactly when and which functions actually switch over to the in-house models.
Self-contained vs. open collaboration
Same goal, different worldview
Security Copilot and Codex Security CLI both aim at "AI for cyber defense," yet they take opposite approaches to how that gets delivered. The key differences:
| Microsoft Security Copilot (in-house defense AI) | OpenAI Codex Security CLI (open source) |
|---|---|
| Bundled into Microsoft 365 E5, single-vendor and self-contained | Released as open source; anyone can adopt or modify it |
| Model details (name, performance figures) undisclosed | Code published under Apache-2.0, open to outside verification |
| Fits organizations already consolidated on Microsoft | Suits teams that want to mix and match vendors |
| Announced amid an active antitrust probe, so friction is possible | Going open source reads as a hedge against lock-in concerns |
Fuller details will come in stages
Microsoft has not disclosed the models' names or concrete performance figures this time. Those specifics will need to wait for future announcements, likely surfacing gradually over the coming weeks to months as Security Copilot update notes.
Recommended action: track detection-logic drift
Detection accuracy and false-positive rates can shift once the in-house models take over. Engineering teams should set up log comparisons of alert volume and content before and after the cutover.
Recommended action: map your Microsoft dependency
Business and procurement teams should inventory how deep their security-operations reliance on Microsoft runs, in step with the antitrust probes' progress. Document the lock-in risk formally, and keep a record of any open-source alternatives considered, so the decision is easy to explain later.
There is plenty to be cautious about. As long as the model internals stay undisclosed, outsiders have little way to independently verify how well the defense AI actually performs. For buyers who favor comparing vendors, strengthening a self-contained model in the middle of an antitrust probe can read as underwhelming rather than decisive. Set next to a transparent option like Codex Security CLI, what Security Copilot owes the market next is verifiable proof of performance.