Mythos leak rattles cyber and crypto

Anthropic’s accidental exposure of internal material about an unreleased AI system dubbed Claude Mythos has jolted investors, sending cybersecurity shares sharply lower and adding to pressure across risk assets, while reviving a deeper debate over whether frontier AI is moving faster than the safeguards meant to contain it. The leaked material, first reported by Fortune on March 26 and 27, described Mythos as Anthropic’s most powerful model yet and said it showed unusually strong performance in coding, reasoning and cybersecurity tasks.

The market reaction was swift on March 27. Palo Alto Networks, CrowdStrike, Zscaler and Fortinet all fell as investors reassessed what a model with stronger vulnerability-finding and exploit-related capabilities could mean for established security vendors. Reports in Barron’s, Bloomberg and MarketWatch said the sell-off was tied directly to fears that a more capable AI system could either outpace current defensive tools or compress the value of some existing software offerings. While the exact percentages varied through the trading day, the direction was unmistakable: security names were hit first and hardest.

What emerged from the leak appears to be more than a routine product disclosure gone wrong. Fortune reported that thousands of unpublished assets were left publicly accessible because of a configuration lapse, and that the exposed material referred both to “Claude Mythos” and to a new model tier called “Capybara”, positioned above Anthropic’s Opus line. The draft language described the system as “by far the most powerful AI model” the company had developed and said it delivered markedly higher scores on software coding, academic reasoning and cybersecurity benchmarks than its previous best model. That combination helps explain why investors treated the episode not as gossip from the AI rumour mill, but as a signal that the technical frontier may be shifting faster than expected.

The sharper question is whether the market has overreached. Some analysts argued the sell-off reflected headline shock more than a settled view of industry economics. Barron’s and MarketWatch both reported that several market watchers see a paradox at work: a model that increases cyber risk could also increase demand for better cyber defence, particularly AI-native monitoring, identity controls and automated response tools. In that reading, Mythos is not simply a threat to security vendors; it is also a possible accelerant for the next spending cycle in security. That view fits a broader theme already visible this year, with investors repeatedly punishing software groups on fears that generative AI could flatten product differentiation before later reconsidering which companies may actually benefit.

Anthropic’s own policies give the story wider significance. The company’s Responsible Scaling Policy, updated in February, says advanced models must be evaluated for catastrophic-risk domains including cyber operations, and that stronger safety and security standards are required when capability thresholds are crossed. Anthropic has publicly framed those rules as a mechanism for delaying deployment or tightening controls if a system’s capabilities outrun available safeguards. Against that backdrop, the leaked description of Mythos matters not only because it unsettled traders, but because it suggests frontier labs may be approaching models that force harder choices over release, restricted access and state-level oversight.

Crypto was also under pressure, though the causal chain there is less clean. Some crypto outlets tied losses in bitcoin and ether to the Mythos revelations, but broader market reporting pointed to other drivers, including options expiry, liquidations and geopolitical risk linked to Middle East tensions. That makes it harder to say the leak alone sent digital assets sharply lower. What can be said with confidence is that the episode fed a wider mood of risk aversion around assets exposed to technology volatility, while also highlighting a more uncomfortable possibility for the sector: models that become better at finding flaws in software and digital infrastructure could, over time, raise the perceived fragility of exchanges, wallets and blockchain-linked applications.