Leak shows security gaps — hiring shift imminent
The Anthropic leak and subsequent demos (including reported zero‑day kernel findings) have forced labs to throttle access and exposed a hiring need: expertise in AI safety, adversarial robustness and system‑level security is moving from niche to core. ( )
A CMS misconfiguration on March 27, 2026 left nearly 3,000 Anthropic internal assets publicly accessible, including draft posts that named an unreleased model called “Claude Mythos” and an internal tier dubbed “Capybara.” (winbuzzer.com) Leaked drafts described Mythos as a “step change” with novel cybersecurity risks, and Anthropic publicly confirmed it was testing the model and removed the exposed materials from public access. (tech.yahoo.com) Anthropic had already demonstrated automated vulnerability discovery in February when Claude Opus 4.6 surfaced more than 500 previously unknown high‑severity zero‑day flaws in open‑source libraries during closed testing. (red.anthropic.com) Following capacity and safety concerns this quarter, Anthropic has limited Claude subscription usage and previously tightened access by blocking third‑party OAuth routes to curb unauthorized integrations. (infoworld.com) Recruiting patterns shifted visibly: OpenAI advertised a Head of Preparedness role with $555,000 in annual cash compensation, and both OpenAI and Anthropic have posted specialist roles targeting weapons‑and‑explosives misuse to address catastrophic‑misuse scenarios. (tech.yahoo.com) Anthropic is also expanding pipeline hiring with a paid 2026 AI Safety Fellows program that offers weekly stipends (about $3,850 USD) for each fellow, while job listings at OpenAI and DeepMind explicitly require adversarial‑robustness, RLHF, and model‑threat‑defense experience and commonly expect a PhD or 4+ years of relevant research experience. (aimpactful.com) OpenAI’s reported plan to nearly double headcount to roughly 8,000 employees by the end of 2026 underscores the scale of demand for safety and systems‑security expertise across frontier labs. (msn.com)
