# Claude Code Leaks, Mythos Escapes, Anthropic Explains **By:** Maria Sukhareva **Published:** 2026-04-12T00:00:00+00:00 **Source:** [Substack](https://msukhareva.substack.com/p/claude-code-leaks-mythos-escapes?triedRedirect=true) --- ### The Most Moral of All Moral Companies Cannot Stop Building Weapons. This Time It Is a Cyber Weapon. Anthropic has been in the news a lot lately, and each time it is not the best image of the most moral of all moral companies. [The company has been known to provide Claude Gov as the main LLM for military use by the US forces.](https://www.cnn.com/2026/02/27/tech/anthropic-pentagon-deadline) The models assisted in operations in Venezuela and Iran. [Claude is a model that reasons over Palantir's ontology and allows the lightning-fast execution of the kill chain.](https://msukhareva.substack.com/p/when-a-thousand-missiles-fly-at-once) Surprisingly, Dario Amodei decided to become the pacifist of the year, picking a fight with the Pentagon and refusing to remove two specific restrictions: the use of Claude for fully autonomous weapons (those that select and engage targets without any human in the loop) and for mass domestic surveillance. [The Pentagon’s position](https://www.axios.com/2026/02/15/claude-pentagon-anthropic-contract-maduro) was that it wanted Claude available for “all lawful use,” with no written carve-outs of any kind. [ ![](https://substackcdn.com/image/fetch/$s_!TGGq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dd939fc-8d5d-4fb1-b5be-2928115c5fae_1536x545.png) ](https://substackcdn.com/image/fetch/$s_!TGGq!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dd939fc-8d5d-4fb1-b5be-2928115c5fae_1536x545.png) Emerging capabilities of Mythos [Suddenly Amodei was hailed as the moralist](https://globalaffairs.org/commentary/analysis/anthropic-pentagon-and-claudes-split-personality) who stands up to the evil military machine. What a coincidence that just one day later, on February 28, the US attacked Iran, with the reasoning assistance of Claude showering the country [with high-precision strikes:](https://www.csis.org/analysis/assessing-air-campaign-after-three-weeks-iran-war-numbers#:~:text=Ukraine%20has%20achieved.-,U.S.%20Strike%20Campaign,around%20$3.5%20million%20per%20shot.) > In the first 24 hours of Operation Epic Fury, U.S. forces [struck](https://www.centcom.mil/MEDIA/IMAGERY/igphoto/2003881835/) over 1,000 targets as they worked from the long-standing U.S. Central Command (CENTCOM) target list. The Israeli Air Force [struck](https://x.com/IDF/status/2029640963044356367?s=20) over 750 additional targets during this time. The US did not achieve its goals in Iran and the regime did not fall, but they managed to scale the error and casualties to [unprecedented measures](https://www.aljazeera.com/news/2026/3/12/who-bombed-the-iranian-girls-school-killing-more-than-170-what-we-know). The strike on Minab’s Shajareh Tayyebeh elementary school killed between 156 and 170 people according to different investigations, most of them children, both boys and girls, with the school serving the families of IRGC naval personnel. Multiple independent investigations point to a combination of AI-assisted targeting and human intelligence failure, with likely use of outdated information that failed to distinguish the school from an adjacent military compound: [ #### Did AI Misidentify the Minab School? ](https://msukhareva.substack.com/p/did-ai-misidentify-the-minab-school) The irony of the situation is that [Sam Altman apparently fell for the stunt](https://fortune.com/2026/03/02/openai-ceo-sam-altman-defends-decision-to-strike-pentagon-deal-amid-backlash-against-the-chatgpt-maker-following-anthropic-blacklisting/) and signed the contract with the Pentagon, becoming the villain of AI, even though those were still Anthropic models being used by Palantir for the strikes. Altman himself later admitted the timing was [“opportunistic and sloppy”](https://www.cnbc.com/2026/03/03/openai-sam-altman-pentagon-deal-amended-surveillance-limits.html), which is about as close to self-awareness as a $300 billion CEO gets. And in a coda that Anthropic’s PR team could not have scripted better, [a federal judge later preliminary ruled](https://www.cnbc.com/2026/03/26/anthropic-pentagon-dod-claude-court-ruling.html) that the Pentagon’s blacklisting of Anthropic was “classic illegal First Amendment retaliation,” calling the government’s reasoning “Orwellian” in a 43-page ruling. Everything was going great for Anthropic: enterprise deals growing, the arch-enemy OpenAI vilified and [losing customers who switched to Claude](https://letsdatascience.com/blog/altman-called-the-pentagon-deal-sloppy-1-5-million-users-had-already-left), and [infrastructure deals locked in with both Microsoft and Google](https://www.cnbc.com/2026/03/06/google-says-anthropic-remains-available-outside-of-defense-projects.html). Life was great for Anthropic. ## Claude Code Leak And then, the problems came from where they were not expected. [A packaging error resulted in the whole Claude Code being published to the npm repository](https://www.bleepingcomputer.com/news/artificial-intelligence/claude-code-source-code-accidentally-leaked-in-npm-package/) - whether by a developer or a vibe coder will remain a mystery. In an attempt to mitigate the leak, Anthropic started taking down repositories and ended up [sweeping more than 8,000 GitHub repos](https://www.firstpost.com/tech/anthropic-accidentally-removes-thousands-of-github-projects-while-removing-leaked-claude-code-13995685.html), including forks of their own public repo, creating more chaos. [ ![](https://substackcdn.com/image/fetch/$s_!Myps!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F606e5a8d-90e4-4171-bbfc-ea35c1130b5b_1498x610.png) ](https://substackcdn.com/image/fetch/$s_!Myps!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F606e5a8d-90e4-4171-bbfc-ea35c1130b5b_1498x610.png) People started questioning the maturity of the product and the general fitness of the company to comply with cybersecurity standards. If they cannot even protect their own product, how can we trust them to comply with complex enterprise processes? The examination of the code confirmed the concerns were justified. [Researchers at Adversa AI found a critical vulnerability](https://www.securityweek.com/critical-vulnerability-in-claude-code-emerges-days-after-source-leak/) where deny rules were silently bypassed for command chains exceeding 50 subcommands, allowing attackers to exfiltrate SSH keys and credentials via a malicious file in a cloned repo. ![](https://substackcdn.com/image/fetch/$s_!bIsR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4a347d94-33bf-4858-af56-b19d827073c1_2572x470.png) [Three further vulnerabilities forming a complete attack chain were found by Oasis Security](https://securitybuzz.com/cybersecurity-news/researchers-say-claude-flaws-could-be-chained-to-silently-exfiltrate-user-data/), and [a zero-click attack in Claude Code’s Chrome](https://www.koi.ai/blog/shadowprompt-how-any-website-could-have-hijacked-anthropic-claude-chrome-extension) extension was found by Koi Security even before the leak. Just a few hours after Claude Code was leaked, [developers started reproducing it.](https://www.scmp.com/tech/tech-trends/article/3348817/anthropics-ai-code-leak-ignites-frenzy-among-chinese-developers) [Sigrid Jin](https://www.aol.com/articles/inside-race-recreate-claude-code-214543644.html), a 25-year-old student at the University of British Columbia, woke up at 4am to his phone blowing up. Two humans, ten OpenClaw accounts, a MacBook Pro, and a few hours later he had rewritten the entire codebase in Python. [The repo hit 75,000 GitHub stars in approximately two hours](https://github.com/ultraworkers/claw-code). A Rust rewrite followed. Anthropic’s DMCA machine had nothing to aim at. It was getting even worse for Anthropic’s cybersecurity image: [hackers started integrating malware into fake leaked Claude Code repositories](https://www.zscaler.com/blogs/security-research/anthropic-claude-code-leak), deploying the Vidar infostealer and GhostSocks malware, optimised for search engines and hitting the first page of Google for anyone searching for the leak. Whether connected or not, around this time [MiniMax published its MMX-CLI](https://cntechpost.com/2026/04/10/minimax-launches-mmx-cli-command-line-tool-expand-ai-business-footprint/), a command-line tool for AI agents with multimodal assistance. It was almost certainly developed independently, but the functional overlap is hard to ignore: both are CLI tools for AI agents, both integrate into coding workflows like Claude Code and OpenClaw, and MMX-CLI explicitly lists Claude Code as a target environment. Connected or not, it illustrates how fast the gap between a proprietary product and what competitors can build is closing. ## Claude Mythos Escape But here another lucky coincidence happened! Apparently, Anthropic is actually the leader in cybersecurity thanks to its [new model Mythos](https://thehackernews.com/2026/04/anthropics-claude-mythos-finds.html). The model is so incredible that it “escaped a sandbox and emailed a researcher”, specifically, Sam Bowman, who [found out while eating a sandwich in a park](https://observer.co.uk/news/business/article/safety-first-puts-anthropic-ahead-in-game-of-ai-spin). Anthropic was so shocked about how good their model is that they in fact decided that [they cannot publish such an impressive and dangerous model.](https://www.nbcnews.com/tech/security/anthropic-project-glasswing-mythos-preview-claude-gets-limited-release-rcna267234) Or maybe it just escaped so far that they still cannot catch it: Anyway, Mythos is still a myth now and is available as a preview for the lucky ones through [Project Glasswing](https://www.anthropic.com/glasswing). What is not a myth though is that [researchers at AISLE tested open-weight models](https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier) on the exact vulnerabilities Anthropic showcased - a finding amplified by HuggingFace CEO Clément Delangue on X. [ ![](https://substackcdn.com/image/fetch/$s_!bJ69!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff991125f-3325-489d-9485-5282c81a5279_1206x1595.png) ](https://substackcdn.com/image/fetch/$s_!bJ69!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff991125f-3325-489d-9485-5282c81a5279_1206x1595.png) The models recovered much of the same analysis. Eight out of eight detected the flagship FreeBSD exploit, including one with only 3.6 billion parameters costing eleven cents per million tokens. Can it be that Mythos is not that special after all? It is special. But it will not stay for long. The honest comparison though is this: Anthropic’s test had Mythos navigate entire codebases autonomously, without being told where to look. The AISLE test had a human first identify the relevant code and hand it to the models for analysis. The models tested were also tiny: 3.6 and 5.1 billion parameters — almost certainly with limited context windows compared to Mythos, which can hold and reason over entire codebases at once. **The capability to detect vulnerabilities in isolated code is already present in small models.** The actual challenge, and what Mythos genuinely demonstrated, is the [needle-in-a-haystack proble](https://github.com/gkamradt/LLMTest_NeedleInAHaystack)m: navigating hundreds of thousands of lines of unfamiliar code, identifying which 50 lines out of 500,000 matter, and writing a working exploit without human help at any step. To credibly claim a breakthrough, or to credibly challenge one, you need to test large open-weight models e.g. MiniMax-2.7, DeepSeek-V3.2, GLM-5.1 - on the same full autonomous pipeline. Nobody has done that yet. If small models already detect on pre-selected code, large open-weight models on a full codebase may tell a very different story. [Just a day later, OpenAI discovered that they also have a tool that is too dangerous to release.](https://sherwood.news/tech/openai-our-new-ai-tool-spud-is-too-dangerous-to-release/) Not [GPT-2](https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction), or [GPT-4](https://www.theguardian.com/technology/2023/mar/17/openai-sam-altman-artificial-intelligence-warning-gpt4), or any other model that scared Altman before. It is a different one, internally known as Spud. [ ![](https://substackcdn.com/image/fetch/$s_!zSXm!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8646d536-074f-4fb2-9978-66a05307bf46_2570x1194.png) ](https://substackcdn.com/image/fetch/$s_!zSXm!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8646d536-074f-4fb2-9978-66a05307bf46_2570x1194.png) ## The mystery of Mythos performance The most mystical part of all this is what is actually the underlying technological, architectural, or algorithmic breakthrough that allows this incredible performance. Because as we know, [there is only one internet](https://the-decoder.com/openai-co-founder-says-ai-is-reaching-peak-data-as-it-hits-the-limits-of-the-internet/), Anthropic has already scraped all there is to scrape, [pirated all there is to pirate,](https://www.npr.org/2025/09/05/nx-s1-5529404/anthropic-settlement-authors-copyright-ai) so what is it that made transformers so much better? That fantastic mythical performance of the Mythos cannot be just god-given. The aggressor is always peace-loving; he would prefer to take over our country unopposed. \-- Carl von Clausewitz, On War Well, no matter how you look at it, most of the gains for transformers come from incremental improvements. [Sparse mixture of experts](https://huggingface.co/blog/moe) allows more efficient use of compute by activating only a fraction of the model’s parameters per token rather than the whole network. Anthropic's own system card for Mythos says the capability gains are "confidently attributable to human research, not AI assistance," but declines to say what that research actually was, citing sensitivity. The specific advances remain undisclosed. Anthropic describes the training process as follows: > Claude Mythos Preview was trained on a proprietary mix of publicly available information from the internet, public and private datasets, and synthetic data generated by other models. Throughout the training process we used several data cleaning and filtering methods, including deduplication and classification. Nothing special here - internet crawl and some model distillation. What we really want to know is what happened during post-training, but they do not tell us: > After the pretraining process, Claude Mythos Preview underwent substantial post-training and fine-tuning, with the goal of making it an assistant whose behavior aligns with the values described in Claude’s constitution. Basically, they said absolutely nothing. Luckily we can read between the lines, and their 244-page system card is surprisingly revealing. In fact, this document, probably unintentionally, allows us to guess with high probability what environment the model was trained in and why its cybersecurity abilities did not simply "emerge." It is highly likely that the most moral of all moral companies, Anthropic, just cannot stop building weapons: this time it is a cyber weapon. **Further in the article:** detailed analysis of the Mythos model card, possible reconstruction of their training environment and tasks, and argumentation for why its hacking capabilities did not just emerge. The new era of cyber weapons is coming and gatekeeping this model is meaningless, harms safety, and is pure PR. **Free subscribers** _get my public analysis._ The stuff [Forbes](https://www.forbes.com/sites/timkeary/2026/03/14/5-highlights-to-watch-at-nvidia-gtc-2026/) and [Die Welt](https://www.welt.de/wirtschaft/article68a41dae99ec9e2fe941d5f6/chatgpt-und-co-das-ende-der-schwarzen-paepstin-doch-auch-die-anti-woke-ki-offenbart-probleme.html) cite. **Paid subscribers** _get what’s too honest to say publicly:_ Vendor comparisons that nobody with partnerships would dare publish. What out of millions of AI announcements is worth paying attention to. The PowerPoint engineering of consulting agencies compared to what actually works. And you can ask me anything about your AI decisions directly. **Founding members** _get all of that plus a 1:1 call with me._ 45 minutes. 16 years in enterprise and research AI, from machine translation to production GenAI systems. Review your use case, validate your startup idea, or pressure-test your AI strategy. You decide how we spend this time.