Project Glasswing found over 10,000 critical vulnerabilities in its first month
Anthropic’s Project Glasswing used Claude Mythos to discover over 10,000 critical vulnerabilities in systemically important software within its first month, establishing AI-assisted vulnerability research at a scale previously impossible. Meanwhile, NVIDIA’s Nemotron-Labs Diffusion models achieved 6x inference speedups through parallel token generation, Microsoft began canceling internal Claude Code licenses in a strategic consolidation toward GitHub Copilot CLI, and a new paper demonstrated that compiling agentic workflows directly into model weights can match orchestrator-based systems at two orders of magnitude less cost.
Security #
Project Glasswing: Anthropic’s AI-assisted vulnerability discovery at scale #
Anthropic / Hacker News (443 points)
Anthropic’s Project Glasswing, a collaborative cybersecurity initiative with approximately 50 partners, has discovered over 10,000 high- or critical-severity vulnerabilities in systemically important software within its first month using Claude Mythos Preview. Partners including Cloudflare (2,000 bugs found) and Mozilla (271 Firefox vulnerabilities) are demonstrating that AI-assisted vulnerability discovery can operate at a scale that shifts the bottleneck from finding vulnerabilities to verifying and patching them. Mythos Preview is the first model to solve the UK AI Security Institute’s cyber ranges end-to-end, but Anthropic has withheld public release due to insufficient safeguards, instead offering defenders Claude Security tools and a Cyber Verification Program.
AI voice reconstruction from crash recordings forces NTSB to restrict access #
TechCrunch / Ars Technica
Users applied AI to spectrogram images of cockpit voice recordings published in NTSB investigation dockets, reconstructing audible speech from what was intended to be unplayable visual data. The NTSB temporarily blocked access to its entire docket system in response. This exploits a gap in the 1994 law that prohibits NTSB from disclosing cockpit audio recordings – the law predates AI capabilities that can reconstruct audio from visual representations of sound, demonstrating how AI transforms previously inert data formats into exploitable sources.
Google Search breaks on “disregard” query due to apparent prompt injection defense #
TechCrunch
Searching for the word “disregard” on Google now returns an empty AI summary block followed by essentially no useful results. The likely cause: “disregard” is a common trigger word in prompt injection attacks (“disregard all previous instructions”), and Google appears to have implemented filtering that treats the word itself as adversarial input. A common English word becoming effectively unsearchable on the world’s largest search engine is a concrete example of how AI safety mechanisms can create collateral damage in production systems.
Autonomous LLM Agents and CTFs: claims of near-human performance do not replicate #
arXiv
This paper revisits recent claims that LLM agents achieve near-human success rates on capture-the-flag security challenges. Across 30 web-based CTF challenges spanning 14 vulnerability classes, different agent architectures instantiated with state-of-the-art models fail to reproduce the reported results. The finding is consistent with a broader pattern: security benchmarks that appear favorable often contain artifacts – static flags, predictable environments, or narrow vulnerability types – that inflate measured performance.
Model Releases #
Gemini 3.5 Flash reaches general availability as speed-optimized agentic model #
Google / Don’t Worry About the Vase (Zvi Mowshowitz)
Google’s Gemini 3.5 Flash is now generally available, positioned as a speed-optimized model for agentic workflows. Independent analysis finds it “likely the best model at its particular speed point,” but with significant caveats: a January 2025 knowledge cutoff, overconfident decision-making, excessive tool calling, and hallucination tendencies. It ranks 9th on the Arena leaderboard and underperforms Claude Opus and GPT-5.5 on intelligence metrics while costing more than previous Flash iterations. The model fills a niche for latency-sensitive agent deployments but falls short of frontier capability claims.
NVIDIA Nemotron-Labs Diffusion: 6x inference speedup through parallel token generation #
NVIDIA / Hugging Face Blog
NVIDIA released Nemotron-Labs Diffusion, a family of language models (3B, 8B, 14B) that generate text by producing multiple tokens in parallel and iteratively refining them, rather than the standard autoregressive one-token-at-a-time approach. The self-speculation mode achieves 6x speedup with comparable accuracy, reaching approximately 865 tokens per second on B200 hardware. Each model supports three generation modes – autoregressive, diffusion, and self-speculation – in a single checkpoint, built by converting pretrained autoregressive models into diffusion models through continued training.
Research & Papers #
Compiling agentic workflows into LLM weights achieves near-frontier quality at 100x less cost #
arXiv
Agent orchestration frameworks (LangGraph, CrewAI, Google ADK, OpenAI Agents SDK, and others) all follow the same pattern: an external orchestrator above the LLM, injecting instructions and routing decisions every turn. This paper demonstrates that for procedural tasks, distilling the orchestration logic directly into model weights through fine-tuning matches the quality of multi-agent systems at two orders of magnitude less inference cost. The implication is direct: for well-defined workflows, the orchestrator is overhead. This does not eliminate the need for orchestration in dynamic, open-ended tasks, but it challenges the default assumption that every agentic workflow requires a framework.
The Log is the Agent: event-sourced architecture for auditable AI agents #
arXiv
Most agent frameworks are built around the language model: conversation loop first, then tools, then logging bolted on for observability. ActiveGraph inverts this arrangement: the append-only event log is the source of truth, the working graph is a deterministic projection of that log, and behaviors are ordinary functions that react to log events. This makes agent execution fully auditable, forkable from any point with modified parameters, and debuggable without the observability afterthought that plagues current frameworks. For teams building agents that need compliance or reproducibility guarantees, the event-sourced pattern offers those properties by construction rather than instrumentation.
TerminalWorld: 1,530 benchmark tasks reverse-engineered from real terminal recordings #
arXiv
TerminalWorld automatically converts 80,870 real terminal recordings into evaluation tasks, producing 1,530 validated benchmarks spanning 18 categories from short operations to 50+ step workflows covering 1,280 unique commands. The data engine approach – deriving benchmarks from observed human behavior rather than constructing synthetic scenarios – addresses a persistent weakness in coding agent evaluation: most benchmarks test what researchers imagine agents should do rather than what developers actually do.
DeltaBox: millisecond-level sandbox checkpoint and rollback for AI agents #
arXiv
AI agent workflows that require state exploration – tree search, reinforcement learning, speculative execution – need rapid checkpoint and rollback of complete sandbox state including files and process memory. Existing mechanisms duplicate the entire state, costing hundreds of milliseconds to seconds per operation and bottlenecking deep search. DeltaBox achieves millisecond-level checkpoint/rollback through incremental state differencing, enabling the high-frequency state management that practical agent training and test-time search require.
Did Google’s AI agents really build an operating system for $916? #
AI as Normal Technology (Narayanan & Kapoor)
Google claimed AI agents autonomously built an operating system from a single prompt for $916. Narayanan and Kapoor identify multiple methodological problems: the “single prompt” was thousands of lines long with unclear human intervention, no analysis checked whether code was copied from existing sources, and Google released no logs or source code for independent verification. As AI capability claims grow more dramatic, the demand for reproducible, independently verifiable evaluation grows with them.
Developer Tools #
OpenAI named Leader in 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents #
OpenAI
Gartner named OpenAI a leader in the inaugural 2026 Magic Quadrant for Enterprise AI Coding Agents, recognizing Codex for innovation and enterprise-scale deployment. The existence of a Gartner Magic Quadrant for AI coding agents signals that the category has matured from experimental tooling to an enterprise procurement decision with established evaluation criteria, and provides the first authoritative third-party ranking in a format that enterprise buyers use to make purchasing decisions.
Kanbots: open source Kanban app with parallel AI agents on every card #
Hacker News (221 points)
Kanbots is an open source desktop application that attaches AI agents to Kanban cards, running them in parallel to work through a project board concurrently. The approach treats project management cards as units of agent work rather than units of human task tracking. The project represents the emerging pattern of embedding agents into existing workflow tools rather than requiring developers to adopt new agent-specific interfaces.
Funding & Business #
Microsoft cancels internal Claude Code licenses, pushes GitHub Copilot CLI #
The Verge / Windows Central / Hacker News (278 points)
Microsoft is canceling Claude Code licenses for its Experiences and Devices division – the teams building Windows, Microsoft 365, Teams, and Surface – with a June 30 deadline to transition to GitHub Copilot CLI. Claude Code had gained widespread adoption among Microsoft developers after access was opened in December 2025. The move is both financial (aligning license costs with Microsoft’s fiscal year) and strategic (consolidating on owned tooling rather than a competitor’s product). For the broader market, it illustrates the tension enterprises face between using the best available AI tools and maintaining vendor alignment.
AI startups inflating ARR metrics with VC awareness and support #
TechCrunch
TechCrunch documents how AI startups are systematically inflating Annual Recurring Revenue figures by substituting Contracted ARR (signed but undeployed contracts) for traditional ARR, sometimes by 70% or more. Additional practices include counting free pilot periods as revenue and extrapolating volatile monthly usage over 12 months. One unnamed enterprise startup claimed $100 million ARR when only a fraction represented paying customers, and investors are reportedly aware of and complicit in these practices.
Threads to Watch #
Agent architecture is being challenged from multiple directions simultaneously. “Compiling Agentic Workflows into LLM Weights” argues the orchestrator layer is unnecessary overhead for procedural tasks. “The Log is the Agent” inverts the standard architecture so the event log, not the model loop, is the source of truth. DeltaBox enables the millisecond-level state management that agent tree search actually requires. These are not incremental improvements to existing frameworks – they are architectural alternatives that question whether the dominant orchestrator-above-model pattern is the right design at all.
Cybersecurity is becoming a primary AI application, not a defensive afterthought. Anthropic’s Glasswing found 10,000+ critical vulnerabilities in a month using capabilities so powerful that the model itself cannot be publicly released. The NTSB incident showed AI transforming inert data into exploitable content. Google’s “disregard” filter demonstrated prompt injection defenses breaking production systems. The CTF reassessment showed prior security agent capabilities were overstated. AI is simultaneously the most powerful vulnerability discovery tool ever deployed and a generator of novel attack surfaces that existing defenses were not designed for.
Independent scrutiny of AI capability claims is intensifying. AI Snake Oil’s dissection of Google’s “$916 OS” claim, the CTF paper’s failure to reproduce reported security agent performance, Zvi’s detailed analysis revealing Gemini 3.5 Flash’s limitations behind the headline metrics, and TechCrunch’s ARR inflation expose all share a pattern: the gap between what AI companies claim and what independent evaluation confirms is receiving more systematic attention.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Stanford HAI — scrape: content_truncated