GenAI PMClaude Mythos Preview
A Claude model preview that Anthropic withheld due to high blast radius. It is cited as an example of a model being held back for deployment-risk reasons.
Key Highlights
- Claude Mythos Preview is described as an Anthropic model preview for autonomous vulnerability finding and exploitation in software systems.
- Anthropic reportedly withheld the model from broad deployment in April 2026 because its blast radius was considered too high.
- The model was still used in restricted settings such as Project Glasswing, where partners reported major gains in vulnerability discovery.
- Its coverage makes it a useful AI PM case study in staged rollout, partner-only access, and defense-in-depth product design.
- Anthropic cited Mythos Preview alongside evidence that model-layer defenses are strong but imperfect without environmental containment.
At a glance
- Category
- Model
Claude Mythos Preview
Overview
Claude Mythos Preview is an Anthropic model preview focused on autonomous software vulnerability discovery and exploitation. Across newsletter mentions, it is described as a highly capable security-oriented Claude variant that helped partners uncover large numbers of critical software flaws, including in major open-source and commercial systems. It also became notable because Anthropic reportedly withheld it from general deployment in April 2026 due to its high potential blast radius.For AI Product Managers, Claude Mythos Preview matters as a concrete case study in frontier-model deployment tradeoffs. It shows that strong capability gains can create product, safety, and governance challenges at the same time: the same model that appears highly valuable for defenders in controlled programs like Project Glasswing was also cited internally as too risky to ship broadly. That makes it a useful reference point for decisions about staged rollout, containment design, permissions, partner-only access, and measuring whether model-layer defenses are sufficient on their own.
Key Developments
- 2026-04-08: Anthropic published a technical report on software vulnerabilities and exploits uncovered in Claude Mythos Preview, detailing flaws, attack vectors, and mitigations.
- 2026-04-18: DeepLearning.AI highlighted Claude Mythos Preview as a model that can autonomously find and exploit critical software vulnerabilities; access was limited to industry partners ahead of any public release.
- 2026-04-23: A Simon Willison item quoted Firefox CTO Bobby Holley saying Claude Mythos Preview helped identify many vulnerabilities in Firefox, reinforcing its practical value for defensive security work.
- 2026-05-23: Anthropic’s initial Project Glasswing update said Claude Mythos Preview was used with about 50 partners and helped surface more than 10,000 high- or critical-severity vulnerabilities in systemically important software; Cloudflare alone reportedly found 2,000 bugs, including 400 high/critical issues.
- 2026-05-29: Anthropic’s Claude Opus 4.8 announcement used Claude Mythos Preview as an alignment reference point, saying Opus 4.8 showed similar misalignment rates while delivering other capability gains.
- 2026-06-03: Anthropic reportedly expanded Project Glasswing, rolling out Claude Mythos Preview to roughly 150 additional organizations across 15+ countries.
- 2026-06-19: Anthropic Engineering described how it contains Claude across products using environmental controls, model-layer controls, and external-content restrictions, noting Claude Mythos Preview was judged too risky to ship in April 2026.
- 2026-06-25: Follow-up coverage reiterated that layered defenses are used across Claude products and that Claude Mythos Preview had been considered too high-risk for release despite strong model-layer defenses.
- 2026-07-19: Anthropic further framed Claude Mythos Preview as an example of a model withheld for deployment-risk reasons, alongside data showing prompt-injection defenses are helpful but imperfect and need overlapping containment layers.
Relevance to AI PMs
1. A model-risk gating example: Claude Mythos Preview is a practical example of a model that may clear capability thresholds but still fail deployment thresholds. AI PMs can use this framing when defining launch criteria that include blast radius, misuse potential, and operational containment readiness.2. A rollout design lesson: The model appears to have been used in restricted partner programs like Project Glasswing rather than broadly shipped. That is a strong template for PMs managing powerful systems: start with limited-access cohorts, high-trust users, explicit use cases, and intensive monitoring before considering wider availability.
3. A defense-in-depth product requirement: Anthropic’s discussion around Mythos Preview underscores that model-layer safeguards alone are not enough. AI PMs should plan for environment controls, permission systems, sandboxing, egress restrictions, and human review workflows as core product requirements, not post-launch add-ons.
Related
- Anthropic: Creator of Claude Mythos Preview and the company that reportedly withheld it from broad deployment due to blast-radius concerns.
- Project Glasswing: The security-focused program through which Claude Mythos Preview was deployed to selected partners for vulnerability discovery.
- Claude / claude.ai / Claude Code / Claude Cowork: The broader Claude product family and deployment surfaces discussed in Anthropic’s containment writeups.
- Claude Opus 4.7: Referenced in Anthropic’s safety discussion as a model with strong but imperfect prompt-injection resistance, helping contextualize why additional containment is necessary.
- Claude Opus 4.8: Compared against Claude Mythos Preview on misalignment rates in Anthropic’s product announcement.
- Cloudflare: A Project Glasswing partner that reportedly found thousands of bugs using the system.
- Firefox / Bobby Holley: Firefox’s CTO was cited describing how Claude Mythos Preview helped identify vulnerabilities in Firefox.
- Dario Amodei, Boris Cherny, Greg Isenberg, Simon Willison, DeepLearning.AI: Individuals and outlets that discussed or amplified reporting around the model and its implications.
Frequently asked questions
What is Claude Mythos Preview?
A Claude model preview that Anthropic withheld due to high blast radius. It is cited as an example of a model being held back for deployment-risk reasons.
What should AI product managers know about Claude Mythos Preview?
Claude Mythos Preview is described as an Anthropic model preview for autonomous vulnerability finding and exploitation in software systems. Anthropic reportedly withheld the model from broad deployment in April 2026 because its blast radius was considered too high. The model was still used in restricted settings such as Project Glasswing, where partners reported major gains in vulnerability discovery.
Explore this topic
Newsletter Mentions (9)
“Anthropic withheld Claude Mythos Preview in April 2026 for having too high a blast radius and reports that model-layer defenses are strong but imperfect—Claude Opus 4.7 holds prompt-injection attack success to about 0.1% on single attempts and roughly 5–6% after 100 adaptive attempts—so overlapping defenses across environment, model, and external content are required.”
#2 📝 Anthropic Engineering How we contain Claude across products - Anthropic deploys three agentic products—claude.ai, Claude Code, and Claude Cowork—using two containment strategies: human-in-the-loop approvals (users approved roughly 93% of permission prompts) and environmental containment (sandboxes, VMs, filesystem and egress controls); Claude Code’s auto mode automates safer approvals and catches roughly 83% of overeager behaviors before they execute. Anthropic withheld Claude Mythos Preview in April 2026 for having too high a blast radius and reports that model-layer defenses are strong but imperfect—Claude Opus 4.7 holds prompt-injection attack success to about 0.1% on single attempts and roughly 5–6% after 100 adaptive attempts—so overlapping defenses across environment, model, and external content are required.
“Layered defenses—environmental sandboxes/VMs/egress controls, model-layer system prompts/classifiers/training, and limiting external content—are used across claude.ai, Claude Code, and Cowork; telemetry shows users approved roughly 93% of permission prompts, Claude Code auto mode blocks about 83% of overeager behaviors before execution, Claude Opus 4.7 holds prompt-injection success to ~0.1% on single attempts (~5–6% after 100 adaptive attempts), and Claude Mythos Preview was judged too high-risk to ship in April 2026.”
The preview model appears in the context of Anthropic's product safety decisions. It was not shipped because of perceived risk.
“Anthropic has deployed Claude across claude.ai, Claude Code, and Claude Cowork while containing blast radius via environment controls (sandboxes, VMs, filesystem/egress limits), model-layer controls (system prompts, classifiers, probes, training), and restricting external-content/tool access, noting Claude Mythos Preview was judged too risky to ship in April 2026.”
📝 Anthropic Engineering How we contain Claude across products - Anthropic has deployed Claude across claude.ai, Claude Code, and Claude Cowork while containing blast radius via environment controls (sandboxes, VMs, filesystem/egress limits), model-layer controls (system prompts, classifiers, probes, training), and restricting external-content/tool access, noting Claude Mythos Preview was judged too risky to ship in April 2026. Telemetry showed users approved ~93% of permission prompts, Claude Code auto mode blocks roughly 83% of overeager behaviors before execution, and Claude Opus 4.7 resists prompt-injection with about 0.1% success on single attempts and ~5–6% after 100 adaptive attempts.
“𝕏 Anthropic expands Project Glasswing, rolling out Claude Mythos Preview to roughly 150 more organizations across 15+ countries.”
#3 𝕏 Anthropic expands Project Glasswing, rolling out Claude Mythos Preview to roughly 150 more organizations across 15+ countries.
“Also, Opus 4.8 is about four times less likely than Opus 4.7 to let code flaws pass unremarked, scored 84% on Online-Mind2Web, was the only model to complete every Super-Agent case end-to-end (beating prior Opus models and GPT-5.5), is the first to break 10% on the Legal Agent all-pass standard, and shows misalignment rates similar to Claude Mythos Preview while Genie users report 61% cheaper token cost versus Opus 4.7 for multimodal reasoning.”
Anthropic releases Claude Opus 4.8 with dynamic workflows #1 📝 Anthropic News Introducing Claude Opus 4.8 - Anthropic released Claude Opus 4.8 today at the same price as Opus 4.7, adding user-selectable effort levels, Claude Code “dynamic workflows,” and a fast mode that runs 2.5× faster and is three times cheaper than on prior models. The company reports broad capability and alignment gains—Opus 4.8 is about four times less likely than Opus 4.7 to let code flaws pass unremarked, scored 84% on Online-Mind2Web, was the only model to complete every Super-Agent case end-to-end (beating prior Opus models and GPT-5.5), is the first to break 10% on the Legal Agent all-pass standard, and shows misalignment rates similar to Claude Mythos Preview while Genie users report 61% cheaper token cost versus Opus 4.7 for multimodal reasoning. Also covered by: @v0 , @There's An AI For That , @There's An AI For That , @Aravind Srinivas , @claire vo 🖤, building @chatprd , @Mike Krieger , @Claude, @Cognition, @Claire Vo , @Dan Shipper , @How I AI Podcast
“Project Glasswing: An initial update - Project Glasswing, launched last month, used Claude Mythos Preview with about 50 partners to surface more than 10,000 high- or critical‑severity vulnerabilities in systemically important software (Cloudflare reported 2,000 bugs, 400 high/critical) and partners say bug‑finding rates increased by over tenfold.”
Anthropic launches Project Glasswing security scanner #1 📝 Anthropic News Project Glasswing: An initial update - Project Glasswing, launched last month, used Claude Mythos Preview with about 50 partners to surface more than 10,000 high- or critical‑severity vulnerabilities in systemically important software (Cloudflare reported 2,000 bugs, 400 high/critical) and partners say bug‑finding rates increased by over tenfold. Anthropic also scanned over 1,000 open‑source projects and estimated 6,202 high/critical issues out of 23,019 total, triaged 1,752 of those (90.6% true positives, 62.4% confirmed high/critical), found concrete exploits including wolfSSL's CVE-2026-5194, and says the new bottleneck is human triage, disclosure, and patching. Also covered by: @Mario Zechner #2 𝕏 Anthropic reports that Claude Mythos Preview uncovers a surge of security vulnerabilities, meaning patching them boosts safety but forces the software industry to scale its processes—as detailed in their initial Project Glasswing update.
“#10 📝 Simon Willison Bobby Holley - A quoted excerpt from Bobby Holley (CTO, Firefox) describing how Anthropic's Claude Mythos Preview helped identify many vulnerabilities in Firefox and expressing optimism that defenders can now decisively win by focusing on security work.”
#10 📝 Simon Willison Bobby Holley - A quoted excerpt from Bobby Holley (CTO, Firefox) describing how Anthropic's Claude Mythos Preview helped identify many vulnerabilities in Firefox and expressing optimism that defenders can now decisively win by focusing on security work.
“DeepLearning.AI highlights Anthropic’s Claude Mythos Preview, an AI model that autonomously finds and exploits critical software vulnerabilities; it’s currently limited to industry partners to uncover and patch flaws before any public release.”
#3 𝕏 DeepLearning.AI highlights Anthropic’s Claude Mythos Preview, an AI model that autonomously finds and exploits critical software vulnerabilities; it’s currently limited to industry partners to uncover and patch flaws before any public release. #4 𝕏 OpenAI research lead Joy Jiao and product lead Yunyun Wang joined Andrew Mayne on the OpenAI Podcast to unveil the new Life Sciences model series for biology, drug discovery, and translational medicine.
“Anthropic published a detailed technical report on software vulnerabilities and exploits uncovered in Claude Mythos Preview, outlining the specific flaws, attack vectors, and mitigation strategies.”
#1 𝕏 Anthropic published a detailed technical report on software vulnerabilities and exploits uncovered in Claude Mythos Preview, outlining the specific flaws, attack vectors, and mitigation strategies. Also covered by: @Anthropic , @Boris Cherny , @Greg Isenberg , @Simon Willison
Related
Anthropic is the company behind Claude Code, an agentic coding tool used for terminal, IDE, and GitHub workflows. The newsletter also references a recent Claude Code release with expanded agent controls and security checks.
Anthropic’s conversational AI model family, referenced here as handling a task better than the speaker since around Opus 4.8. It is being discussed in terms of agent capability and workflow delegation.
A prominent AI blogger and commentator referenced in connection with an article on token reselling and fraud. He is cited as the source of the newsletter item discussing the marketplace and API-key abuse.
An AI education company sharing Andrew Ng’s views on agent infrastructure requirements. The newsletter cites it as the channel for his argument that agents need more infrastructure, not just better models.
Entrepreneur and product creator known for AI-building demos and distribution-oriented product thinking. In this newsletter, he showcases practical products built with Gemini 4 Argon.
A Claude-associated leader or practitioner cited for a structured approach to agent work. He describes giving agents goals, metrics, and time bounds, then requesting progress artifacts.
A Claude model version referenced for its prompt-injection resistance metrics. It serves as a benchmark example of model-layer defenses being strong but not sufficient on their own.
Cloudflare provides web infrastructure, bot protection, and edge services that are increasingly used for AI agent monetization and control.
Stay updated on Claude Mythos Preview
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free