Capability bounding as product architecture: what Claude Mythos and Project Glasswing actually mean
On April 7, 2026, Anthropic announced its most capable model and, in the same announcement, said it did not plan to make it generally available.
TL;DR
Anthropic’s Claude Mythos Preview found thousands of zero-days, including bugs in every major OS and browser, scored 83.1% on the CyberGym vulnerability reproduction benchmark, and autonomously found and exploited a 17-year-old FreeBSD remote code execution bug (CVE-2026-4747). Instead of releasing it generally, Anthropic created Project Glasswing: up to $100M in usage credits for a dozen launch partners and over 40 more organizations that maintain critical software, for defensive security work. This is not alignment research and it is not plain access control. It is capability bounding: deployment constraints, usage monitoring and output safeguards that restrict what a model does in practice without reducing what it can do in principle. The contrarian read: capability bounding also creates premium pricing power ($25/$125 per million tokens) and a competitive moat that safety alone does not explain. For the threat model these capabilities change, see The AI threat model every security team is missing.

What did Claude Mythos Preview actually find?
The numbers need to land before the architecture discussion makes sense.
Mythos Preview is a general-purpose frontier model that turns out to be extremely effective at computer security. The figures below come from Anthropic’s Glasswing announcement and its red team write-up. On CyberGym vulnerability reproduction, it scored 83.1% versus Opus 4.6’s 66.6%. On SWE-bench Verified, it hit 93.9%. On the OSS-Fuzz corpus, it produced 595 crashes at tiers 1 and 2, where Opus and Sonnet 4.6 reached tier 1 in 150 to 175 cases and tier 2 about 100 times, and it achieved full control flow hijack on ten separate, fully patched targets.
The raw vulnerability count tells part of the story: thousands of zero-days, including bugs in every major operating system and every major web browser. At announcement, Anthropic said over 99% of the vulnerabilities it had found were not yet patched. But specific findings tell the rest:
- FreeBSD NFS (17 years old, CVE-2026-4747): Mythos fully autonomously identified and exploited a remote code execution bug in the NFS server’s RPCSEC_GSS code, starting as an unauthenticated user anywhere on the internet and ending with root. The ROP chain was too long for one request, so the exploit split it across six sequential RPC requests, finishing by appending the attacker’s key to root’s authorized_keys file.
- OpenBSD TCP/SACK (27 years old, now patched): A signed integer overflow that lets an attacker crash any OpenBSD host that responds over TCP. The specific run that found it cost under $50, though Anthropic notes that number only makes sense with hindsight; across a thousand runs the total was under $20,000.
- FFmpeg H.264 codec (16 years old): A bug in a line of code that automated testing tools had hit five million times without ever catching the problem.
- Linux kernel privilege escalation chains: Nearly a dozen examples chaining two, three and sometimes four vulnerabilities to root, including one that used a vulnerability to bypass KASLR.
- Firefox JavaScript engine: Against already-known Firefox 147 bugs in a test harness, Opus 4.6 developed working exploits twice out of several hundred attempts. Mythos Preview developed working exploits 181 times and achieved register control on 29 more.
Expert penetration testers told Anthropic that Mythos Preview wrote exploits in hours that would have taken them weeks to develop. This is not incremental improvement. It is a functional difference in kind.
What is capability bounding and how does it differ from alignment?
The industry language around this is a mess. Three distinct mechanisms exist for controlling what an AI system does, and most commentary mashes all three together.
Alignment operates on the model itself. Through training, RLHF, or constitutional methods, you change the model’s weights so it refuses harmful requests or follows specific behavioral norms. The capability is reduced at the parameter level. The model genuinely cannot (or strongly resists) producing certain outputs.
Access control operates on the API layer. Rate limiting, authentication, role-based permissions, content filtering on inputs and outputs. The model retains full capability, but the serving infrastructure gates who can reach it and what passes through.
Capability bounding operates at the deployment layer. The model retains its full weights and full capability. Access control alone does not define it. Instead, a combination of deployment constraints (who gets access, for what use cases), usage monitoring (what they do with it, with mandatory reporting), and behavioral conditioning (the model is steered toward defensive patterns through system prompts and fine-tuning that does not reduce raw capability) creates an envelope around the model’s practical use.
graph TB
subgraph "Three Control Mechanisms"
A["Alignment<br/><i>Change the model</i>"] --> W["Modify weights via RLHF,<br/>constitutional AI, training"]
B["Access Control<br/><i>Gate the API</i>"] --> X["Auth, rate limits,<br/>content filters"]
C["Capability Bounding<br/><i>Constrain deployment</i>"] --> Y["Use-case restrictions,<br/>monitoring, reporting,<br/>behavioral conditioning"]
end
subgraph "What Changes"
W --> W1["Model capability reduced"]
X --> X1["Model access restricted"]
Y --> Y1["Model use directed"]
end
style A fill:#e74c3c,color:#fff
style B fill:#f39c12,color:#fff
style C fill:#2ecc71,color:#fff
The distinction matters because alignment has costs. Reducing a model’s capability to prevent misuse also reduces its capability for legitimate use. Access control is binary: you are either in or out, with limited granularity. Capability bounding attempts to preserve the full capability for approved purposes while restricting the deployment envelope.
Project Glasswing is the clearest public implementation of capability bounding at the frontier. Mythos Preview is not aligned away from finding exploits. That is its purpose. It is not merely behind an API gate. The participating organizations have API access. It is bounded: access is for defensive security work, Anthropic commits publicly to vulnerability details with cryptographic hashes before disclosure, and it is building output safeguards plus a Cyber Verification Program for security professionals whose legitimate work those safeguards would block.
How does Project Glasswing enforce capability bounding in practice?
The enforcement architecture has four layers, each reinforcing the others.
Layer 1: Partner vetting. Access requires organizational approval. Launch partners (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks) are not random API customers. They are organizations with existing security infrastructure, incident response teams, and reputational incentives to comply. The additional 40+ organizations build or maintain critical software infrastructure.
Layer 2: Terms of access. Access is granted for defensive security work, and partners agree to share information and best practices with each other where they can. The detailed partner agreements are not public. Whatever they say, a named participant carries legal and reputational exposure that an anonymous API customer does not. Anthropic’s own commitment is public: it will report what it has learned within 90 days.
Layer 3: Commitments and audit. For vulnerabilities it cannot yet disclose, Anthropic published SHA-3 hashes of the details, to be replaced with full write-ups once fixes ship. That creates a tamper-evident record of what was found and when. It is an audit chain for vulnerability disclosure, not just logging of API calls.
Layer 4: Behavioral steering and output safeguards. Glasswing work is framed around defensive workflows: scanning for vulnerabilities, generating patches, producing disclosure reports. Anthropic also describes safeguards that “detect and block the model’s most dangerous outputs”, which it plans to refine on an upcoming Claude Opus model before relying on them more broadly. This is lighter than alignment (the weights are not changed to remove exploit capability) but heavier than a content filter.
graph LR
subgraph "Capability Bounding Stack"
L1["Partner Vetting<br/>Organizational approval"] --> L2["Terms of Access<br/>Defensive purpose, info sharing"]
L2 --> L3["Commitments and Audit<br/>Hashed disclosures, public report"]
L3 --> L4["Behavioral Steering<br/>Defensive system config,<br/>output safeguards"]
end
subgraph "Enforcement Properties"
L1 -.-> P1["Reputational incentive"]
L2 -.-> P2["Legal and reputational exposure"]
L3 -.-> P3["Tamper evidence"]
L4 -.-> P4["Operational direction"]
end
style L1 fill:#3498db,color:#fff
style L2 fill:#2980b9,color:#fff
style L3 fill:#1abc9c,color:#fff
style L4 fill:#16a085,color:#fff
Each layer alone is insufficient. A vetted partner could still misuse the model. Contractual terms are enforceable only after violation. Hashed commitments prove what was known when, but do not stop misuse. Behavioral steering can be circumvented by sophisticated users. Together, they create enough friction that the expected cost of misuse exceeds the expected benefit for any partner organization.
Why did Anthropic choose capability bounding over alignment for Mythos?
Because alignment would destroy the product.
Mythos Preview’s value proposition is finding and exploiting vulnerabilities. An aligned model that refuses to produce exploit code is useless for defensive security research. You cannot scan a codebase for vulnerabilities if the model has been trained to avoid discussing vulnerabilities. The defensive and offensive capabilities are the same capabilities applied to different targets.
This is the core tension that makes capability bounding architecturally interesting: the capability you want to restrict for safety is the same capability you want to preserve for value. Alignment cannot thread this needle. Access control is too coarse. Capability bounding attempts to solve it at the deployment layer.
Anthropic plans to “launch new safeguards with an upcoming Claude Opus model, allowing us to improve and refine them with a model that does not pose the same level of risk as Mythos Preview.” The strategy is explicit: use a less capable model to iterate on safety mechanisms, then apply those mechanisms to the frontier model. This is capability bounding as an iterative engineering process, not a one-time policy decision.
Is capability bounding actually a competitive moat disguised as safety?
Here is the contrarian read that the safety narrative obscures.
Claude Mythos Preview is priced at $25 per million input tokens and $125 per million output tokens. For reference, Opus 4.6 sits at $5/$25. That is a 5x premium on both input and output. The pricing alone signals that this is not charity work.
Restricted access creates scarcity. Scarcity creates premium pricing power. Premium pricing funds the infrastructure to maintain restricted access. The flywheel is self-reinforcing: build the most capable model in a high-stakes domain, demonstrate the capabilities loudly (thousands of zero-days, 27-year-old bugs, 83.1% on CyberGym), restrict access to vetted organizations, charge a premium justified by the capability gap, then use the safety argument to prevent competitors from undercutting through open release.
OpenAI had already moved in the same direction. GPT-5.3-Codex (February 2026) was the first launch OpenAI treated as High capability in cybersecurity under its Preparedness Framework, and it launched a “Trusted Access for Cyber” program on day one: the same restricted-access pattern, two months before Glasswing. GPT-5.4 Thinking (March 2026) was its first general-purpose model shipped with mitigations for High cybersecurity capability.
The safety rationale is genuine. Mythos found vulnerabilities that survived decades of expert human review and millions of automated test executions. Releasing that capability to anyone with an API key would be reckless. But the business rationale is equally real, and not everyone accepted the restriction. OpenAI’s Boaz Barak, quoted by Zvi Mowshowitz, wrote: “I encourage Anthropic to release Mythos, even if it’s a version that over refuses on cyber tasks.” Mowshowitz himself concluded that the restriction “was necessary”. The risk on the other side is diffusion: if other labs reach similar capability and release it with fewer controls, the responsible actor gets the moat and the less careful actor gets the market.
Simon Willison put it more simply: “I think the security risks really are credible here, and having extra time for trusted teams to get ahead of them is a reasonable trade-off.” He is right. But trade-offs that happen to generate premium revenue tend to stick around longer than trade-offs that do not.
What precedents exist for capability bounding in AI?
Capability bounding is not entirely new, but Mythos is the first implementation at this scale with this level of public architecture.
GPT-2 staged release (2019): OpenAI initially withheld the full GPT-2 model over misuse concerns and released it in stages over the following months. That was bounding by delay rather than by deployment envelope.
GPT-4 staged release (2023): OpenAI released GPT-4 without full details of training data, architecture, or compute. This was information bounding, not capability bounding. The model itself was generally available.
DALL-E content restrictions (2022-present): Image generation models with filters preventing certain content categories. This is closer to alignment (the model is modified) plus access control (content filters at the API layer) than capability bounding.
Nuclear technology export controls: The closest physical analogy. Centrifuge technology is not destroyed. It is bounded: export-controlled, restricted to approved states, monitored by the IAEA, subject to inspection regimes. The capability exists. The deployment envelope is constrained. The enforcement relies on layered mechanisms (treaties, inspections, sanctions) rather than any single control.
Pharmaceutical scheduling: Fentanyl is not banned from production. It is capability-bounded: manufactured by approved facilities, distributed through controlled channels, prescribed by licensed practitioners, monitored through prescription databases. The capability (pain relief) and the risk (addiction, death) are the same molecule.
Same logic every time: when capability and risk are inseparable, you bound the deployment rather than destroying the capability.
What does capability bounding mean for security teams?
If you run a security program, capability bounding changes your planning in three ways.
First, vulnerability discovery velocity is now asymmetric. Mythos can find vulnerabilities in hours that took human researchers weeks. If your organization is a Glasswing partner, you have access to this velocity for defense. If you are not, attackers may gain similar capabilities before you do. The question is not whether AI-discovered zero-days will affect your infrastructure. It is whether you will find them first. For a framework to evaluate this, see Building an AI security program: from ad-hoc to governance.
Second, your threat model needs a new actor. Autonomous AI vulnerability discovery is no longer theoretical. Mythos Preview found a 17-year-old RCE in FreeBSD with no human involvement after the initial request. Your threat model should now include “AI-automated vulnerability discovery and exploitation” as a distinct capability tier, separate from human-assisted and fully manual. For a structured approach to this, see AI security assessments: evaluating systems you did not build.
Third, patching cadence assumptions are obsolete. If an AI model can find thousands of zero-days in weeks, the assumption that you have months between vulnerability discovery and exploitation collapses. Anthropic’s own disclosure policy gives maintainers up to 90 days, plus up to 45 more, after it reports a bug. Windows like that assume attackers are not finding the same bugs in parallel. Defenders need to plan for compressed timelines.
What happens when capability bounding fails?
Capability bounding is not permanent. Three failure modes are predictable.
Capability diffusion. The exploit generation capabilities emerged in Mythos “as a downstream consequence of general improvements in code, reasoning, and autonomy.” They were not explicitly trained. This means any sufficiently capable frontier model may develop similar capabilities as a side effect of general improvement. Anthropic cannot bound capabilities that emerge in competitor models.
Enforcement decay. Contracts work until someone breaks them. A partner organization could be compromised. An employee could exfiltrate techniques. Coordinated disclosure windows create a known delay that sophisticated adversaries could exploit. Cryptographic hashing proves what was known when, but does not prevent the knowledge from spreading.
Economic pressure. The up-to-$100M in credits is a subsidy. At $25/$125 per million tokens, sustained scanning of large codebases is expensive. When the credits run out, partner organizations face a choice: pay premium rates for defensive scanning, or reduce coverage. Economic pressure will eventually thin the defensive network that capability bounding created.
None of these failure modes hit tomorrow. But they are structural, and they compound. Capability bounding buys time, not permanence. The question is whether the time it buys is enough to develop better mechanisms, which is exactly what Anthropic says it is doing with the upcoming Claude Opus safeguard iteration.
Where does this leave the industry?
Capability bounding is now an established pattern. OpenAI and Anthropic both use it for their strongest cyber models. The next frontier model with dual-use capabilities will face the same architectural decision: align it (and lose the capability), release it (and accept the risk), or bound it (and capture the premium).
From a safety perspective, this is pragmatic in a way that alignment research is not. You do not need to solve the alignment problem (still unsolved) or trust access control alone (insufficient). You build a deployment architecture that preserves capability for legitimate use while making misuse expensive and traceable.
From a business perspective, pay attention. Restricted access to the most capable model in a high-stakes domain, premium pricing, and a safety narrative that shields you from competitive pressure to open up? That is a business model, not just a safety policy. It works because the safety argument is real, but the revenue is real too.
The honest assessment: capability bounding works while frontier cybersecurity capabilities exist at only two or three labs. When those capabilities diffuse (and they will, because they emerge from general capability improvement), the deployment constraints that define capability bounding will need to evolve into something more durable. What that looks like is the open question.
Key takeaways
- Capability bounding is distinct from alignment and access control. It operates at the deployment layer, preserving full model capability while constraining its practical use through partner vetting, contractual obligations, technical monitoring, and behavioral steering.
- Mythos Preview’s findings are real and unprecedented. Thousands of zero-days, including bugs in every major OS and browser, 83.1% on CyberGym vulnerability reproduction, and autonomous exploitation of vulnerabilities that survived decades of human review.
- Project Glasswing is the clearest public implementation of capability bounding at the frontier: up to $100M in credits, a dozen launch partners plus over 40 more organizations, access for defensive work, hashed disclosure commitments.
- The safety argument and the business argument are the same argument. Restricted access creates both responsible deployment and premium pricing power.
- Capability bounding buys time, not permanence. As capabilities diffuse to other labs, the enforcement mechanisms must evolve.
Update, November 2026
At the time of writing, Anthropic still had not made Mythos-class models generally available. Its October 2026 Cyber Verification Program update said Glasswing partners uncovered at least 129,000 verified software vulnerabilities between April and July 2026, more than 33,000 of them rated critical or high severity, and expanded verified access to Mythos models through three tiers of trusted access. The pattern this post describes has become the default for the most capable cyber models rather than a one-off.