Microsoft is preparing to launch Project Perception, an AI cybersecurity platform reported to be arriving as early as this month, and its most interesting feature is not that it finds software vulnerabilities. It is how the system decides which AI model to use for each step. Rather than pushing every task through one powerful model, Project Perception routes work across models from Microsoft, OpenAI, and Anthropic, choosing the cheapest one that can do each job. That design choice is the story, because it reflects a broader shift in how serious AI products are being built. This piece looks at what Project Perception is, why the routing matters economically, and what is still unknown.
Microsoft
OpenAI
AnthropicWhat Project Perception is reported to do
According to reporting, Project Perception is a platform that looks across a company's code, cloud infrastructure, and endpoints, identifies exploitable weaknesses, explains their impact, and proposes concrete fixes. In other words, it aims to cover the full loop from finding a vulnerability to recommending how to close it, rather than just flagging issues for a human to triage.
The stated ambition is continuity. The goal Microsoft is reported to be chasing is to make AI-driven vulnerability finding and fixing cheap enough to run nonstop, rather than as an occasional scan. That framing is the key to understanding the architecture, because always-on only works if each unit of work is inexpensive.
Why routing across models is the whole point
Running a frontier model is expensive, and a security platform that scans continuously would generate an enormous volume of small tasks. If every one of those tasks, from listing assets to parsing a log line to triaging a common issue, went through the most capable and most expensive model, the economics would not support running it around the clock. The cost would force the tool back into occasional, batch-style scans, which is exactly what continuous monitoring is meant to replace.
Project Perception's reported answer is an orchestration layer that matches each job to the best-fit model. A low-cost model handles the high-volume, low-difficulty work: inventory checks, log parsing, first-pass triage of common vulnerability types. When the system needs to reason through a complex exploit chain, interpret an authentication flow, or write a remediation plan that touches production systems, it escalates to a frontier model. The expensive reasoning is reserved for the small share of tasks that genuinely require it.
Routing each job to the cheapest model that can do it
The orchestration idea behind cost-controlled AI security: reserve expensive frontier reasoning for the small share of jobs that need it, and let cheaper models absorb the high-volume work.
Illustrative tiers, not measured prices. The economic point is that running a single frontier model on every step is what makes continuous, always-on scanning unaffordable.
Always-on AI is an economics problem before it is a capability problem. Routing is how you make the math work.
The pattern is bigger than security
What makes Project Perception worth attention beyond cybersecurity is that its central design decision, do not send everything to the biggest model, is becoming a default in well-built AI systems generally. The reasoning is the same everywhere. Model capability has become abundant and increasingly interchangeable at the top, while cost still varies by more than an order of magnitude between a small model and a frontier one. When capability is plentiful and price is the binding constraint, the smart move is to treat models as a portfolio and route by task, not to standardise on a single one.
Notice, too, what the multi-provider sourcing implies. A product this visible drawing on Microsoft, OpenAI, and Anthropic models at once is a signal that even the largest platform companies are designing for optionality rather than committing a flagship product to one model family. It hedges against any single provider's price changes, rate limits, outages, or a rival simply shipping a better model next quarter. The same logic that makes continuous scanning affordable also makes the product resilient to a fast-moving model market.
The open questions
Several important details remain unconfirmed, and they matter for judging the product rather than the idea. Microsoft has not publicly announced Project Perception's availability, pricing, supported workloads, or which customers will be eligible, and much of what is known comes from reporting rather than official documentation. It is also entering a contested space: the tool is being positioned against Anthropic's Mythos in AI security, so it will be measured against a well-regarded incumbent rather than launching into open field.
There are harder questions specific to routing, as well. A system that decides which model handles a security-critical judgement is only as good as that routing logic, and a misroute, sending a subtle exploit chain to a model too small to catch it, is a failure mode that a single-model system does not have. Multi-model orchestration trades a cost problem for a routing-quality problem, and how well Project Perception manages that trade is exactly what real-world use will reveal.
The through-line for anyone building on models
Strip Project Perception down to its principle and it is a lesson that applies far beyond security. In a market where a new frontier-class model arrives almost every week and prices keep falling, wiring a product to a single model is both more expensive and more fragile than routing across several and picking the right one per job. That is true for a security scanner running millions of checks, and it is true for an everyday workflow choosing between a quick answer and a hard one.
The mechanics differ by scale, but the posture is the same: keep the workflow fixed and let the model be a choice. A model-agnostic workspace such as Metir AI is built on that principle for general work, routing each task to a fitting model rather than committing everything to one, which is the same bet Project Perception is reported to be making inside cybersecurity.
The bigger picture
Project Perception is notable less for what it detects than for how it is built. Its multi-model orchestration is a concrete, high-profile example of a design principle that is quietly becoming standard: capability is abundant, cost is the constraint, and routing across models is how you reconcile the two. The vulnerability-finding features will be judged against strong incumbents, and the key operational details are still unconfirmed. But the architecture is a clear signal of where serious AI products are heading, toward treating models as an interchangeable portfolio rather than a single dependency.
Sources:
- Microsoft's Project Perception Could Challenge Anthropic's Mythos in AI Security | TechRepublic
- Microsoft's Project Perception Aims to Make AI Vulnerability Fixing Cheap Enough to Run Nonstop | Windows News
- Microsoft Prepares Project Perception to Compete With Anthropic Mythos | Windows Report
- Microsoft to launch Project Perception AI cybersecurity platform this month | NewsBytes
Image credits
Header image: Building 92 on Microsoft's Redmond campus, Washington, by Jiaqian AirplaneFan via Wikimedia Commons, licensed under CC BY 3.0. In-body portrait of Satya Nadella by Brian Smale for Microsoft via Wikimedia Commons, licensed under CC BY-SA 4.0.