Is the Mythos Threat a Myth?
This week, Anthropic announced Claude Mythos Preview, a frontier model it says has reached a level of coding capability that allows it to “surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” The company further stated that it had already identified thousands of zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old bug in OpenBSD and a 16-year-old flaw in FFmpeg.
Citing the risk of bad actors exploiting these same capabilities, Anthropic is limiting access to a curated group of major tech and financial firms, including Amazon, Apple, Microsoft, Google, JPMorganChase, and roughly 40 others, under a new initiative called Project Glasswing, backed by $100 million in usage credits. According to some reports, the announcement prompted Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell to convene an emergency meeting with the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo, and Goldman Sachs to discuss the systemic risks the model poses to the financial system.
However, the panicked reaction may warrant skepticism. AI companies have a documented pattern of warning loudly about the dangers of their own technology, a dynamic critics say is motivated by the desire for regulatory capture. The specific cybersecurity claims around Mythos have also gotten pushback. Researcher Boyan Milanov called Anthropic’s claims “largely unsubstantiated” and said there was no evidence Mythos meaningfully outperforms prior models at vulnerability discovery rather than exploitation. Another commentator pointed out that Mythos's full-exploit rate drops to 4.4% once two repeated bugs are removed, implying four distinct bugs total rather than the fifty the model card implies. So while it’s no doubt true that AI will pose security issues as it continues to improve, the hype around Mythos may have more to do with its marketing than with the underlying reality.
Intel Scores with Terafab
This week, Intel announced it will join SpaceX and Tesla’s Terafab project, a new semiconductor initiative targeting one terawatt per year of compute to power AI, robotics, satellites, and SpaceX’s proposed space data centers. Intel’s contribution fills an obvious gap: SpaceX and Tesla have no experience in chip fabrication, and building a leading-edge fab typically requires years and more than $20 billion.
Back when we published Building an American TSMC, we argued that Intel Foundry’s survival is the only realistic path for the US to rebuild domestic semiconductor manufacturing capacity, and that its central problem was a lack of anchor customers willing to commit to American-made chips. As we pointed out, hyperscalers and large technology companies are “incentivized to invest in semiconductor capacity independent of Taiwan, both in terms of geopolitical risk and domestic regulatory scrutiny”. The Terafab partnership is exactly the kind of demand signal Intel can take advantage of. Whether Terafab ultimately delivers on its ambitions, its clearest near-term effect is giving Intel Foundry what it has been missing: two high-profile anchor customers.
Meta’s Claude Problem
Three stories about Meta made headlines this week. First, a graphic from The Information noting that Meta employees used 60.2 trillion tokens in 30 days. Second, a story around Meta’s internal “Claudenomics” leaderboard ended as quickly as it began, with the company shutting down the leaderboard after reports of intentionally inefficient usage. Third, Meta announced Meta Spark, the company’s first proprietary model developed under Alexandr Wang’s Meta Superintelligence Lab.
On the surface, it isn’t likely that Meta’s use of Anthropic tools was a huge deal. It has been pointed out that many of the 60.2 trillion tokens consumed were likely input tokens contextualizing coding tasks. This means that, while significant, extrapolating the token count to conclude that Meta accounts for one-third of Anthropic’s revenue is incorrect. While the company hasn’t made public comment about shutting down Claudenomics, a reasonable read could be that (1) a system that incentivizes inefficient token usage is a bad system (2) Meta wants their employees to use the new model, even though it’s worse than other SOTA models on coding, (3) Meta doesn’t want the launch of the new model to be overshadowed.
It’s possible, however, that these stories are connected. Followers of Meta’s AI progress have hypothesized that the company intentionally encouraged the usage of Anthropic’s models to collect data used to distill reasoning traces from Claude models. This isn’t a crazy idea; Bloomberg reported that Muse Spark was training using “third-party open-source models, including Qwen from the Chinese tech giant Alibaba Group Holding Ltd., as well as ones from OpenAI and Google.” A Meta spokesperson said, “Like others across the industry, Meta uses techniques like distillation with strict safeguards in place to learn from openly available AI models and improve our own.”
What We’re Reading
In addition to Anthropic, OpenAI is apparently also finalizing an invite-only cybersecurity product offering access to its most capable new models for defensive purposes. Both it and Anthropic have now adopted restricted rollout strategies for frontier AI amid growing industry alarm that autonomous hacking capabilities have arrived.
OpenAI, Anthropic, and Google began sharing threat intelligence through the Frontier Model Forum to detect adversarial distillation attempts by Chinese competitors, a practice in which Chinese labs systematically query U.S. models to train cheaper copycat versions, costing American AI labs an estimated billions of dollars annually.
Anthropic is reportedly in the early stages of exploring in-house AI chip design as a strategic response to industry-wide semiconductor shortages, while currently relying on a mix of Google TPUs and Amazon chips, mirroring similar efforts by Meta and OpenAI.
Perplexity’s annualized recurring revenue jumped 50% to more than $450 million in March after the company launched a new AI agent tool and shifted to usage-based pricing, as it reported more than 100 million monthly active users across its search and agent products.
Chinese automaker BYD unveiled an EV charging system capable of recharging a battery as fast as filling a conventional gas tank, according to a widely circulated post on Twitter.
The Trump administration directed the FAA to develop regulatory frameworks for eVTOL aircraft (colloquially called “flying cars”) following executive orders that created an Integration Pilot Program now underway across 26 states with participants including Joby, Archer, and Wisk, with public operations expected to begin by summer 2026.
A Gallup survey of over 1.5K Americans aged 14 to 29 found that the share of Gen Z who say AI makes them excited fell from 36% in 2025 to 22% in 2026, while those who say it makes them angry rose from 22% to 31%, with Gallup researchers attributing the shift in part to older Gen Zers’ concerns about AI eliminating entry-level jobs.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!