Anthropic inadvertently left about 3,000 blog-related files reachable through an open content system, including drafts and materials on an unreleased model called Mythos. The company says the material was provisional and unrelated to its AI systems or customer data. This analysis separates what is confirmed from what remains unverified.
🔑 Key takeaways
- About 3,000 files tied to Anthropic’s blog were accessible without authentication before access was secured.
- The Mythos model, reportedly positioned above the Opus family, is described in drafts as far stronger in coding, reasoning and cybersecurity.
- More than 80% of code merged into Anthropic’s codebase was authored by Claude as of May 2026.
- METR says Claude Mythos Preview could work autonomously for at least 16 hours.
An Exposure of Unpublished Content
Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge, found that roughly 3,000 files linked to Anthropic’s blog sat in an open data cache. They included drafts, documents, images and presentation materials that had never been published on the company’s public sites. Fortune alerted Anthropic on a Thursday, when access was locked down. The exact number of inaccessible URLs has not been disclosed.
| Item | Detail |
|---|---|
| Accessible files | About 3,000 (Cambridge estimate) |
| Date of lockdown | A Thursday, after Fortune’s alert |
| Exact URL count | Not disclosed |
| Model named in drafts | “Claude Mythos” or “Mythos” |
Before the lockdown, Anthropic stored blog articles, images and documents in a system reachable without authentication. The system could return files to anyone who formulated the right requests, even when publication had not been decided. Some resources also had addresses that opened directly in a browser.
Cybersecurity professionals say the system defaulted to making uploaded items public, including logos, illustrations and research articles, unless access was explicitly restricted. Anthropic reportedly failed to apply that restriction to documents that were never meant to be public. A single misconfigured default was enough to turn a working repository into an unintended showcase.
Mythos, Positioned Above Opus
The official confirmation Fortune received concerns the existence of a new general-purpose model that Anthropic describes as a “step change”. The company says it is being developed and tested with early-access customers, and that it delivers substantially higher performance than its predecessors in reasoning, coding and cybersecurity. An internal summary calls it the company’s “most capable model to date”.
“A problem with one of our external content management tools made temporary content accessible.”
Anthropic spokesperson
More specific details come from the drafts reviewed by journalists, not from a standalone public statement. SmarterX reports that the documents projected significantly higher scores than Claude Opus 4.6 on software engineering, academic reasoning and cybersecurity. One line attributed to the drafts says Mythos is “currently far ahead of any other AI model in cyber capabilities” and foreshadows a wave of models able to exploit vulnerabilities well beyond what defenders can match.
That cyber claim should be read as material from internal documents, not as a detailed official position. The same dataset also referenced an invitation-only retreat in the United Kingdom for executives of large European companies, which Dario Amodei was expected to attend. A spokesperson described it as “part of an ongoing series of events” the company has hosted over the past year.
Measured Internal Acceleration
Anthropic’s own research on its development process helps frame the Mythos Preview claims. The company says its engineers now ship eight times as much code per quarter as they did during 2021-2025. In the second quarter of 2026, the typical engineer merged eight times as much code per day as in 2024. Anthropic cautions that lines of code are an imperfect productivity measure and likely overstate the real gain.
| Metric | Value | Reading |
|---|---|---|
| Code merged per engineer (Q2 2026 vs 2024) | 8x per day | Imperfect measure, gain likely overstated |
| Share of code authored by Claude (May 2026) | Over 80% | Low single digits before February 2025 |
| Median self-estimated speedup (March 2026 poll, 130 staff) | 4x | Plausible, but probably too high |
| Speedup on optimization experiments | 52x (April 2026) | Claude Opus 4 reached 3x in May 2025 |
| Longest autonomous task (METR) | 16 hours or more | Upper limit of what METR can measure |
The length of autonomously handled tasks is also growing. Anthropic describes a shift from reliable task length doubling roughly every seven months to doubling about every four months. In March 2024, Claude Opus 3 completed software tasks that take humans about four minutes. A year later, Claude Sonnet 3.7 handled tasks of about an hour and a half, and Claude Opus 4.6 later reached 12-hour tasks. If the trend holds, Anthropic expects multi-day tasks to become feasible in 2026 and multi-week tasks in 2027.
On open-ended engineering work, Claude’s success rate reached 76% in May 2026, up 50 percentage points in six months. In April, Anthropic’s engineering teams used Claude to ship more than 800 fixes that cut one category of API errors by a factor of 1,000. The supervising engineer estimated that a human would have needed four years. An automated Claude reviewer could also have caught about one third of the bugs behind past claude.ai incidents before production.
Limits That Remain Significant
In research, Claude can match or outperform skilled humans when the objective and experimental method are well specified. Anthropic nonetheless acknowledges large gaps in its ability to judge which engineering or research goals deserve pursuit. The company frames this as partial delegation rather than full recursive self-improvement: an AI system cannot yet design and fully develop its own successor, and Anthropic says that outcome is neither inevitable nor imminent.
Precedents, Market Reaction and the Security Debate
This is not the first accidental exposure of pre-release material. Fortune recalls that in 2018 Apple left the names of iPhones announced hours before launch visible in a public sitemap. Game companies such as Epic Games and Nintendo have previously exposed images, game assets or media ahead of release through content delivery networks or test servers. Google has left internal documentation at public addresses, and Tesla-related data has been reachable through misconfigured third-party servers.
Discoverability is a newer variable. AI coding tools can automate crawling, pattern detection and correlation of publicly available assets. Scripts generated with Claude Code or Codex could scan entire datasets for file names a human might overlook. Fortune nonetheless reports that no AI tool was involved in Anthropic’s configuration error.
According to SmarterX, confirmation of a model with exceptional cyber capabilities was followed by a sharp selloff in security stocks: CrowdStrike, Palo Alto Networks and Zscaler reportedly fell about 6% each, Okta and Netskope more than 7%, and Tenable 9%. Fortune and other reproductions of its report do not independently corroborate these figures.
“More powerful models worry me. They speak precisely to reasoning, coding and cybersecurity. We already knew every model was improving on these fronts. What concerns me is that people are not prepared for what already exists, and that we are very close to much higher-level models.”
Paul Roetzer, founder and CEO, SmarterX
Conclusion: Basic Controls Matter More Than Ever
The company characterizes the episode as a configuration failure rather than a breach of its core infrastructure, AI systems or customer data. Still, it shows how a poorly governed default setting can expose strategic material just as development speed rises. Three scenarios remain open: rapid tightening of access controls and configuration reviews; wider use of automated tools that surface what companies believe is hidden; or a market reassessment of cybersecurity vendors if the claimed capabilities are confirmed publicly.
Sources
- Fortune, exclusive report on the Anthropic leak (March 2026)
- SmarterX, analysis of the Mythos model
- Yahoo Tech, republication of the Fortune report
- FutureSkills Academy, incident analysis
- Anthropic, research on AI-assisted AI development
- Gary Marcus, critical analysis of Anthropic
Cet article est publié à titre informatif et éducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute décision.
