On June 2, two of the biggest names in AI shipped vulnerability-discovery updates within hours of each other. Anthropic expanded Project Glasswing — its joint cybersecurity initiative built on Claude Mythos Preview — to roughly 150 more organizations across more than 15 countries, bringing total access to around 200 partners spanning power, water, communications, healthcare, and hardware. Microsoft, at Build 2026, took its multi-model agentic scanning harness (codename MDASH) out of preview, wired its findings directly into Microsoft Defender, and posted 96.55% on the public CyberGym leaderboard — up from 88.45% at the system's May 12 launch.
Same day. Same problem. Two opposite architectures.
The architectural disagreement is the story
The press coverage on June 2 read both announcements as "AI for cybersecurity, expanded." That's the surface. The substance is what each vendor chose to put behind the API.
Anthropic put one model behind it. Claude Mythos Preview — a single frontier model so capable Anthropic won't make it generally available until the cybersecurity safeguards are tight. Since the program launched, Anthropic and its initial 50-partner cohort have reported more than 10,000 high- or critical-severity vulnerabilities surfaced by Mythos across systemically important open-source software. The thesis: one extraordinarily good model, kept in a closed program, finds bugs at a rate that used to require legions of human researchers.
Microsoft put a council behind it. The MDASH harness orchestrates more than 100 specialized agents across what Microsoft's security blog calls "a configurable panel of models, ranging from state-of-the-art (SOTA) models as the heavy reasoners, to more cost-effective models for high-volume operations." Auditor agents propose candidate vulnerabilities. Debater agents validate or kill them. A separate SOTA model does adversarial review. Microsoft's own framing on the design choice: the panel "allows it to trade speed, recall, and cost, and minimize dependency on any specific model."
Read that last clause again. Minimize dependency on any specific model. That is the council thesis. Microsoft Security shipped it as product architecture.
Stakes
CyberGym is not a toy benchmark. UC Berkeley's Sunblaze Lab built it from 1,507 real-world vulnerability reproduction tasks drawn from 188 open-source projects in Google's OSS-Fuzz continuous-fuzzing campaign — actual exploit reproductions, not synthetic puzzles. The benchmark was published at ICLR 2026. A 96.55% on that leaderboard means MDASH is reproducing real CVEs at a rate that did not exist a few months ago. The May 12 number was 88.45%; roughly three weeks of harness tuning and the Microsoft Defender pipeline integration added eight points.
Mythos sits at the other end of the credibility line. A 10,000-plus high-or-critical-severity scan validated by a 50-partner initial cohort, expanding now to operators of critical infrastructure in more than 15 countries. The framing Anthropic keeps returning to is the bottleneck reversal: progress in software security used to be limited by how quickly vulnerabilities could be found. Now it's limited by how quickly they can be verified, disclosed, and patched.
Both are real. Both work. They work by structurally different means.
What each architecture buys you
Single-frontier-model architecture (Mythos): peak per-query reasoning. One model that has internalized enough of the world's code to spot patterns a human reviewer would miss. The bet is that capability concentration buys correctness — and that the model itself, audited by Anthropic and the partner orgs, is the apparatus.
Multi-model panel architecture (MDASH): error correction by structural disagreement. A configurable panel means no single model's blind spot becomes the system's blind spot. The debater agents catch the auditor agents' false positives. The adversarial SOTA reviewer catches what slipped past both. The costs — coordination overhead, latency, more inference calls — are paid in exchange for the property Microsoft named explicitly: independence from any one model's failure mode.
Neither is universally right. They sit at different points on the same trade-off curve. The choice the buyer is being asked to make is the architectural question, not the vendor question: do you trust one extraordinary reviewer, or do you trust a structured panel?
For high-severity CVE discovery on OSS infrastructure — where false negatives are catastrophic and false positives are merely expensive — both vendors are betting on architectures they've validated empirically. That they shipped opposing validated architectures on the same day is the part the coverage missed.
What this means above cybersecurity
The cybersecurity use case is the cleanest possible test bed because the ground truth is unambiguous: either the bug exists or it doesn't. Most questions a council gets asked are not like that. They are pricing decisions, strategy calls, hiring choices, contract reviews — questions where "is this answer right?" cannot be checked by running a fuzzer.
That's exactly where the architectural divergence matters more, not less. When the ground truth is unverifiable, the cost of betting on one very capable reviewer is that you cannot independently audit its answer. A panel of differently-trained models gives you the only thing structural that resembles a check: a disagreement signal you can read directly. The auditor agent flagged X. The debater agent rejected it. That argument log is the verification step you don't get from a single reviewer no matter how good it is.
Anthropic is betting that Mythos is so good the disagreement signal isn't needed. Microsoft is betting that no matter how good any single model gets, the disagreement signal is structurally what you want a security system to produce.
The interesting part of June 2 wasn't that two AI vendors shipped cybersecurity updates. It was that they made the architectural disagreement public and shipped it as product. One review from the smartest possible reviewer, or a structured argument between models that disagree by design. That's the choice above cybersecurity too, for every decision where wrong is expensive and the answer can't be fuzzed.
If you're trying to decide between one model's opinion and a council's position, that distinction is worth running on your own questions before you commit.
Try it free — no signup. shingik.ai