Microsoft Project Perception: AI Security Tool vs Mythos

The AI security market has a new entrant, and it is aiming at price rather than prestige. According to reporting from The Information and follow-on coverage this week, Microsoft is preparing to launch Project Perception, an AI tool that scans an organization’s code, cloud infrastructure, and endpoints to find exploitable weaknesses and propose fixes — and its central pitch is that it will run far cheaper than Anthropic’s Mythos, the incumbent that has defined the category. Microsoft reportedly plans to introduce the product this month. As of July 18 the company has not published pricing, availability, supported workloads, or customer eligibility, so treat the specifics below as reported rather than confirmed.

The strategy is familiar from Microsoft’s broader AI playbook: let a rival prove out a premium category, then undercut it with a lower-cost, more deeply integrated alternative distributed through channels the rival cannot match.

What Project Perception is reported to do

Project Perception is described as a tool that looks across a company’s codebase, cloud environment, and endpoint fleet, identifies vulnerabilities an attacker could exploit, explains the impact of each weakness, and proposes concrete fixes. That end-to-end loop — find, explain, remediate — is the same shape as the AI security agents that have emerged over the past year, and it leans on the AI agent pattern of a model planning and executing multi-step work rather than answering a single query.

The distinguishing design choice is that Project Perception is multi-model. Rather than routing everything through one frontier model, it reportedly combines models from Anthropic, OpenAI, and Microsoft, sending each individual query to whichever model best fits the task. A cheap classification step might go to a small, fast model; a hard exploit-analysis step might go to a frontier model. The point of the routing is economic: by matching model cost to task difficulty, Microsoft aims to make continuous scanning cheap enough to run nonstop rather than as an occasional audit.

That is a meaningful shift. Security scanning has historically been bounded by cost — you scan on a schedule because scanning everything, all the time, with a top-tier model is prohibitively expensive. Drive the per-scan cost down far enough and the cadence can flip from periodic to continuous, which changes what the tool can catch.

The Mythos comparison

Project Perception is positioned explicitly against Anthropic’s Mythos, the AI security product that established the premium end of this market. The gap the reporting emphasizes is price. Mythos carries an estimated API cost roughly 100% higher than Anthropic’s own Opus model and about 82% higher than GPT, two of the most expensive publicly available models — a reflection of how much frontier-model inference a thorough security pass consumes. Microsoft’s multi-model routing is the mechanism intended to close that gap: use the expensive models only where they are needed and cheaper ones everywhere else.

Microsoft’s structural advantage extends beyond cost. The company already sits inside enterprise environments through Windows, Azure, and Microsoft Defender, which gives it distribution and telemetry a standalone security product has to earn from scratch. Notably, Microsoft has already been building an AI-driven vulnerability-discovery pipeline internally — the multi-model agentic scanning harness (MDASH) it credited for the record volume of its July 2026 Patch Tuesday, which fixed a company-record number of CVEs. Reporting suggests Project Perception is effectively the productization of that internal capability, packaging the same multi-model bug-hunting approach for customers to run against their own systems. If accurate, it means Microsoft is not building the tool from a standing start — it is shipping something it has already been running against its own products.

The multi-model catch

The same architecture that makes Project Perception cheap also makes it complicated to trust. “Multi-model” is a cost-saving decision to Microsoft’s engineers; to a chief information security officer, it is a data-governance problem with several parties involved. Every query routed to an external model from Anthropic or OpenAI means an organization’s security-sensitive data — source code, infrastructure configuration, descriptions of live vulnerabilities — is leaving Microsoft’s boundary and touching a third party’s systems.

Security buyers will therefore press on the questions that determine whether the tool is adoptable at all: isolation between tenants, customer-data retention, model-training boundaries (is my code used to train anyone’s model?), access controls, auditing, regional processing for data-residency rules, and exactly how external models are invoked. A vulnerability report is among the most sensitive artifacts an organization holds — it is, by definition, a map of how to break in. Feeding that map through a routing layer that dispatches pieces of it to multiple vendors is precisely the kind of flow a mature zero-trust security program is built to scrutinize.

There is a second, subtler risk. A tool that ingests untrusted inputs — code comments, log data, configuration files — and feeds them to models that then take actions is exposed to prompt injection, where malicious text embedded in the scanned material tries to manipulate the model’s behavior. Any AI security scanner has to defend against being turned against the organization it is meant to protect, and a multi-vendor pipeline widens the surface where that defense has to hold.

Why Microsoft is moving now

The timing is not incidental. AI-assisted vulnerability discovery has crossed from research demo to production reality this year, visible in the sheer volume of bugs now surfacing — Microsoft’s own record patch month is the clearest signal. Once AI can find flaws faster than humans can, the bottleneck shifts to triage and remediation at scale, and that is exactly the workload Project Perception targets. The company that supplies the tooling for that loop stands to sit at the center of enterprise security spending.

It also fits Microsoft’s competitive posture toward its own AI partners. The company has been increasingly willing to compete with the model providers it also depends on — recently training its salesforce to position its in-house AI against OpenAI, Anthropic, and Google on cost, integration, and enterprise security. Project Perception is the same thesis expressed as a product: use rivals’ models where useful, wrap them in Microsoft’s distribution and pricing, and compete on the economics. It slots into the broader discipline of software supply-chain security, where continuous, automated scanning is becoming the expected baseline rather than a premium add-on.

What it means

Project Perception is a bet that AI security will be won on cost and distribution, not model prestige — and that bet has consequences for everyone in the category.

Who wins. Microsoft, if the economics hold: it can undercut a premium incumbent using rivals’ own models while keeping customers inside its ecosystem. Enterprises win optionality and, potentially, cheaper continuous scanning than a single-frontier-model tool can offer. Anthropic and OpenAI win as suppliers — their models get invoked either way — even as they lose on the product layer where Mythos and its peers command premium margins.

Who’s exposed. Anthropic’s Mythos faces direct margin pressure the moment a “good enough, much cheaper, already-in-your-stack” alternative ships. The premium-security thesis holds only if buyers conclude that a single trusted model is worth paying multiples more than a routed multi-model pipeline — a judgment CISOs will make on governance grounds as much as on detection quality.

What to watch. First, whether Microsoft actually ships this month and on what terms — pricing, data-handling commitments, and which workloads are covered will decide how seriously security teams engage. Second, the governance answers: the multi-model design is the whole cost advantage, but it is also the whole objection, and Microsoft’s contractual guarantees on data retention, training boundaries, and residency will determine adoption more than any benchmark. Third, detection parity: cheaper only wins if it finds the bugs that matter, and independent evidence that a routed pipeline matches a frontier-only tool on real vulnerabilities is still missing. Until Microsoft puts Project Perception in front of customers, the pitch is compelling and the proof is pending.