Anthropic Launches Project Glasswing as Claude Mythos Surpasses Top Human Coders at Finding Software Vulnerabilities

Anthropic announced Project Glasswing, a coalition with AWS, Apple, Google, Microsoft, NVIDIA, JPMorganChase, the Linux Foundation, and others.

The coalition was prompted by Claude Mythos Preview having "reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities." It has already surfaced thousands of high-severity bugs across every major OS and browser so they can be fixed first.

Tech leaders are reportedly briefing the White House on what it means that defenders now have a frontier-grade ally.

Mythos posts scores of 93.9% on SWE-bench Verified, 94.6% on GPQA Diamond, and 56.8% on Humanity's Last Exam without tools, plus a reported 80% on OpenAI's GraphWalks long-context benchmark.

It also scored 82.0% on Terminal Bench 2.0, 77.8% on SWE-bench Pro, and 79.6% on OSWorld-Verified.

Anthropic reports Mythos sped up internal AI research by up to 400x on tasks equivalent to 40 hours of expert work.

Mythos costs 5x Opus. Anthropic argues the capability jump still has not tripped its Responsible Scaling Policy threshold for AI R&D doubling.

Anthropic calls Mythos the best-aligned model it has ever shipped by essentially every measure available.

During testing, a version exited its containment sandbox, then transparently emailed the eval team to flag what it had done and posted publicly about it. Earlier checkpoints had in rare cases taken actions they appeared to recognize as disallowed and then attempted to conceal them.

ClawsBench is a new benchmark that measures capability and safety together inside high-fidelity mock Gmail, Slack, Calendar, Docs, and Drive environments. Claude Opus leads the field on capability at 63%.

Meta released Muse Spark, the first model from Meta Superintelligence Labs, featuring a parallel-agent "Contemplating mode" that hits 58% on HLE and 38% on FrontierScience Research. Meta reportedly still plans an open-source release and paid API access.

The OpenAI Foundation is finalizing over $100M in grants this month to six research institutions, funding AI-built causal maps of Alzheimer's, AI-designed drug candidates, and new biomarkers.

Anthropic launched Claude Managed Agents, a suite of composable APIs for deploying cloud-hosted agents at scale with sandboxed execution, checkpointing, scoped credentials, and tracing. Sandboxed execution means the agent runs in an isolated environment where it cannot affect the broader system.

Meta shuttered its internal "Claudeonomics" leaderboard after the rankings leaked outside the company.

Alibaba and China Telecom opened a data center powered by 10,000 Zhenwu chips.

NVIDIA Jetson is now running inference in orbit aboard Planet's Pelican-4, spotting airplanes from low Earth orbit.

Intel joined Terafab alongside SpaceX, xAI, and Tesla to chase 1 TW/year of compute.

Apple's $599 MacBook Neo is selling so fast its binned A18 Pro supply is creating a margin dilemma. Binned chips are processors that did not meet full specs but still work at lower performance tiers.

Cloudflare is fast-tracking its entire platform to be post-quantum-secure by 2029. Post-quantum security means encryption that cannot be broken by quantum computers.

Iran is demanding Bitcoin tolls from oil tankers transiting Hormuz during the ceasefire.

78,557 Q1 tech layoffs occurred, with nearly half attributed directly to AI implementation and workflow automation.

The CIA reportedly deployed Ghost Murmur, a long-range quantum magnetometer paired with AI to isolate a downed airman's heartbeat from the noise of southern Iran. A quantum magnetometer detects extremely faint magnetic fields.