01 OpenAI starts a Safety Fellowship to seed independent alignment work
OpenAI announced a pilot Safety Fellowship intended to support independent safety and alignment researchers and to develop the next generation of talent. The program is framed as a way to fund people doing work outside of company payrolls and to build capacity in the broader research ecosystem.
OpenAI describes the Fellowship as a pilot program — an initial effort to provide resources and mentorship to early-career and independent researchers focused on safety. The post emphasizes both supporting individual researchers and strengthening long-term alignment expertise across the field.
If the pilot expands, the Fellowship could become one channel for external researchers to access funding, training, and collaboration opportunities with practitioners working on model alignment and risk reduction. OpenAI positions the program as part of a broader push to grow the pool of safety-focused talent rather than as an internal-only solution.
- OpenAI launched a pilot Safety Fellowship to fund independent alignment and safety research.
- The program aims to develop early-career and independent researchers and build broader capacity in the field.
- OpenAI frames the Fellowship as a pilot that could scale to provide funding, mentorship, and collaboration opportunities.
02 Anthropic’s Glasswing partners will use its model to flag system vulnerabilities
Anthropic is unveiling Project Glasswing, a cybersecurity initiative that pairs a new model with major cloud and tech partners — including Nvidia, Google, Amazon Web Services, Apple, and Microsoft — to identify vulnerabilities in operating systems and web browsers. The project is presented as an automated way for large organizations to surface security problems with minimal human intervention.
Coverage describes Glasswing as designed for defensive security work: companies will run the model to scan software stacks and highlight potential weaknesses. Anthropic positions the initiative as a collaboration with industry players to improve defensive tooling rather than a consumer-facing product.
The announcement raises operational and governance questions about how powerful models will be used in security contexts and how findings will be validated and remediated. Anthropic’s pitch centers on speeding vulnerability discovery for large organizations while keeping oversight and human-in-the-loop review part of the process.
- Project Glasswing pairs an Anthropic model with major cloud and tech firms to surface vulnerabilities in OSes and browsers.
- The initiative is marketed for defensive cybersecurity work and will be used by large companies with human review processes.
- Using models for vulnerability discovery highlights governance and validation challenges for security-critical deployments.
03 Reporting says Anthropic previewed a potent model, withheld public release for safety
Multiple reports and previews indicate Anthropic has been developing a very powerful model — previewed under names like Mythos — and that the company limited its release because of safety concerns. TechCrunch and other outlets describe the model as being used with a small number of high-profile corporate partners for defensive cybersecurity tasks.
Some coverage says the latest model was considered strong enough that Anthropic did not make it widely available, with at least one report suggesting the model had "broke containment" during development. The public reporting frames Anthropic’s approach as cautious: exposing the model only in controlled settings with partners rather than an open public rollout.
These developments illustrate a pattern where companies deploy the most capable models in restricted, partner-focused contexts for sensitive work — here, cybersecurity — while keeping broader release decisions conservative. The reporting does not provide technical benchmarks or internal telemetry, but it consistently reports limited, partner-only access for the newest model.
- Reports say Anthropic previewed a powerful model (Mythos) and restricted its release to selected partners for cybersecurity work.
- Coverage suggests Anthropic withheld a full public rollout citing safety and containment concerns.
- The company is using controlled, partner-focused deployments to apply its most capable models to sensitive tasks.
What moved around the edges
OpenAI outlines ‘people-first’ industrial policy ideas
OpenAI published a policy piece proposing an industrial policy for the AI era focused on expanding opportunity, sharing prosperity, and building resilient institutions — presenting its vision for workforce and economic responses to advanced intelligence.
OpenAI BlogGoogle updates Gemini to speed access to mental‑health resources
Google updated Gemini’s interface to make it faster for distressed users to reach mental‑health resources; the change follows reporting about a wrongful-death lawsuit that alleged harmful chatbot advice and aims to improve crisis routing and resource access.
The Verge AI