Anthropic’s $100M Cyber Defense, the Agent Infrastructure Stack, and Models That Learn from Mistakes
Managed agent architectures, self-evolving agent groups, on-the-job learning, and multimodal RAG as a library call.
Anthropic launched Project Glasswing with 11 partners — AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — committing $100M in usage credits and $4M to open-source security organizations. The initiative centers on Claude Mythos Preview, an unreleased frontier model with advanced vulnerability discovery capabilities.
Mythos Preview discovered a 27-year-old vulnerability in OpenBSD that allowed remote crash just by connecting, a 16-year-old FFmpeg bug in a line of code hit 5 million times by automated tools, and chained Linux kernel vulnerabilities for autonomous privilege escalation.
Meta Muse Spark: 10x Efficiency Over Llama 4
Meta released Muse Spark, the first model from Meta Superintelligence Labs — a natively multimodal reasoning model with tool-use, visual chain of thought, and multi-agent orchestration. The headline: more than 10x compute efficiency over Llama 4 Maverick to reach the same capability level.
“Contemplating mode” runs multiple agents reasoning in parallel, scoring 58% on Humanity’s Last Exam. Meta also released an updated Advanced AI Scaling Framework, shifting from rules-based to principle-based safeguards — training the model on reasoning behind restrictions rather than just prohibitions.
Anthropic Managed Agents: Separating State, Logic, and Execution
Anthropic published its managed agents architecture — a hosted service virtualizing agents into three decoupled components: session (append-only log), harness (agent loop), and sandbox (execution environment). The result: agents respond ~60% faster on average, with worst-case latency cut by over 90%.
Previously, each agent ran in a single container — if that container failed, everything was lost. The new architecture decomposes agents into interchangeable stateless components. Session logs live outside the context window, enabling flexible context management. Multiple stateless harnesses scale independently, connecting to sandboxes only when needed. Security improves because credentials never reach the sandbox where generated code runs.
The design enables customer VPC access without network peering, multi-environment coordination (containers, phones, emulators), and workflows that exceed context limits. Anthropic notes model behaviors change over time — “context anxiety” disappeared in newer models — so the architecture keeps harness logic separate from model behavior.
Group-Evolving Agents: 71% SWE-bench Without Human Design
UC Santa Barbara researchers introduced Group-Evolving Agents (GEA), treating groups of agents — not individuals — as the evolutionary unit. The approach achieves 71.0% on SWE-bench Verified (vs 56.7% for the best self-evolving baseline) and 88.3% on Polyglot (vs 68.3%), matching human-designed frameworks through meta-learning alone.
The mechanism is experience sharing across evolutionary branches. GEA pools tool-use traces, patches, and strategies into a shared pool that guides adaptive improvements. It self-adjusts strategy based on problem structure, recovers from bugs in 1.4 iterations versus 5 for tree-structured evolution, and transfers gains across foundation models — suggesting the improvements are architectural, not model-specific.
ALTK-Evolve: Teaching Agents to Learn on the Job
IBM Research released ALTK-Evolve, a long-term memory system that converts raw agent interaction traces into reusable guidelines. Most AI agents repeat mistakes instead of generalizing — a recent MIT study found 95% of AI pilots fail because agents don’t adapt.
The system captures trajectories through observability layers, extracts patterns, consolidates and prunes weak rules, then retrieves the top 5 relevant guidelines just-in-time per task — keeping context lean rather than bloating prompts. Results: +8.9% aggregate on AppWorld, with a 74% relative increase on hard tasks (19.1% to 33.3%). Critically, improvements generalize to unseen tasks, suggesting principle learning rather than memorization. Integration ranges from Claude Code hooks to a single Python import to full MCP protocol.
Multimodal Sentence Transformers: Cross-Modal RAG in a Library Call
HuggingFace‘s Sentence Transformers v5.4 added native multimodal support — encoding texts, images, audio, and video through the same familiar API widely adopted for text embeddings. The release includes embedding models (Qwen3-VL 2B/8B, Nvidia Nemotron) and reranker models for cross-modal retrieve-and-rerank pipelines.
Cross-modal search becomes model.encode() and model.similarity(). The retrieve-and-rerank pattern works across modalities: fast embedding retrieval on the full corpus, then precise cross-encoder reranking on top-k results. VLM models need ~8GB VRAM; CLIP models run on CPU. For RAG teams, this lowers multimodal retrieval from “build a custom pipeline” to “upgrade a dependency.” The modality gap (cross-modal similarities are lower than within-modal) is a known tradeoff, but relative ordering is preserved.
The open question
AI agents are accumulating experience, sharing discoveries across groups, and running on infrastructure that separates what they know from where they execute. The same autonomy that lets agents learn and adapt also lets them find vulnerabilities humans missed for decades. The production stack for autonomous AI is being built in the open. Capability and infrastructure are advancing together — whether the safeguards keep pace is the open question.