Tech News

Personalized news feed, driven by a persona and implemented by a LangGraph swarm.

Anthropic launches new Claude Fable and Mythos 5.2 models

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, claiming improvements in coding, reasoning, and vision over the 5.1 series. Fable 5.2 is designed for complex software engineering, agent orchestration, and long-running projects, and has solved two open theoretical physics problems during testing. Mythos 5.2 is limited to vetted organizations, while Fable 5.2 is available today.

Why it matters

For a solutions architect, these models' demonstrated ability to orchestrate agent teams and solve real engineering and scientific problems signals a shift toward AI as a practical, autonomous collaborator in complex SDLC and infrastructure tasks.

AI/ML / arstechnica.com

Claude Code is accidentally exposing secrets all over GitHub

Apiiro researchers discovered over 100,000 Claude Code settings files on public GitHub, with ~1,100 from 63 organizations containing secrets like database passwords and API tokens. The leak occurs when users approve shell commands with 'Yes, and don’t ask again,' storing the full command—including embedded secrets—in .claude/settings.local.json. While Anthropic adds this file to the global .gitignore, it isn’t repo-level protected, so cloning or reinitializing a repo can expose it; Claude Code may also commit changes without user input.

AI/ML / dev.to

Take a Claude Agent SDK agent from install to deploy with Agents CLI

Google's Agents CLI v1.12.0+ now provides a first-party extension for Claude Agent SDK agents, wrapping them with the same create, run, eval, and deploy commands previously exclusive to ADK agents. The extension handles session persistence via Cloud SQL, tracing, metrics, and eval grading using the same format as ADK, so teams skip building operational infrastructure. A single `uvx google-agents-cli setup` command scaffolds a project targeting Cloud Run, with a plain Python agent file using `ClaudeAgentOptions` and MCP servers, and eval scores 5/5 on default test cases.

AI/ML / techmeme.com

Meituan releases weights for LongCat-2.5, a 1.6T-parameter MoE model for long-horizon agentic tasks, under the MIT license, after unveiling it on November 1 (Carl Franzen/VentureBeat)

Meituan has released the weights for LongCat-2.5, a massive 1.6 trillion-parameter Mixture-of-Experts (MoE) model designed for long-horizon agentic tasks, under the permissive MIT license. This follows the model's initial unveiling on November 1, with the weights release on November 9. The model appears to be targeted at complex, multi-step AI agent workflows.

AI/ML / dev.to

Introducing NovaDesk AI: Building a Stateful Hotel Concierge with LangGraph

NovaDesk AI is an open-source, stateful hotel concierge built with LangGraph, FastAPI, and Chainlit, demonstrating multi-step agentic orchestration beyond basic chatbots. It integrates advanced RAG via HyDE and query rewriting, state and memory management, and human-in-the-loop (HITL) security for cancellations. The system routes natural language intents to SQLite database transactions for policy Q&A, room bookings, and room service, with strict domain guardrails to prevent off-topic queries.

Security / arstechnica.com

Vulnerability that lets hackers SSH into Linux devices is under active exploitation

Hackers are actively exploiting CVE-2026-123158, a critical vulnerability in OpenSSH when paired with the musl library, allowing unauthenticated SSH access to privileged Linux accounts in seconds. The flaw, dubbed *sshole, requires specific conditions: no PAM, passwordless accounts with '*' in /etc/shadow, and SHA256/512 password hashing. Attackers trigger it by sending passwords over 256 characters, causing musl's hashing function to return '*' and bypass authentication. Greynoise reports group GT10 conducting mature intrusions with AI-generated backdoors, while watchTowr and Previdian confirm botnet and cryptocurrency-related exploitation. Alpine Linux systems are most at risk.

General / notesbylex.com

Fable 5.2 one-shotted my nephew's dream game

Fable 5.2, an early release of Anthropic's Claude agent, autonomously built a complete multi-level beat 'em up game from a child's hand-drawn designs in a single overnight session, replacing most of the developer's prior code with polished cutscenes, destructible environments, and controller support. The agent used external APIs (OpenAI ChatGPT Images 2.5 for sprites, ElevenLabs for sound) and committed code incrementally, demonstrating a leap in autonomous game development capability. This marks a shift from earlier agentic tools that required months of iterative prompting to produce playable results.

Languages / infoq.com

Effect 4.0 Ships: Dependency-Free Rewrite with 5x Smaller Bundles and Long-Term Support

Effect 4.0, a from-scratch rewrite of the TypeScript framework, ships with a 5x smaller bundle (7.1 kB vs 35.6 kB), 86% less memory usage, and 6.4x higher throughput. The core now has zero runtime dependencies after removing `pg` and MessagePack, and consolidated modules from `@effect/platform`, `@effect/rpc`, and `@effect/cluster` into the main package. A new LTS policy guarantees bug fixes until September 2029, while the team also added explicit stability tags to all APIs and migrated HTTP, SQL, RPC, and AI spans to OpenTelemetry semantic conventions.

General / dev.to

Traefik 3.8's tap middleware triples request latency even with an instant local receiver

Traefik 3.8.0-rc.1's tap middleware, designed for audit logging by copying requests/responses to a service, triples baseline latency even with an instant local receiver (2.6ms vs 0.88ms). Each request incurs two synchronous calls (request and response), so a 50ms receiver delay adds ~100ms, not 50ms. With rejectOnRecordError: true, an unreachable tap service causes 500 errors for all requests; with rejectOnRecordError: false, failures only appear in logs without built-in metrics, requiring custom alerting on middlewareType=Tap log lines.

AI/ML / infoq.com

Article: LLM Reliability Layers Under Test: Accuracy, Yield, and the Limits of Verification

Testing six reliability layers (deterministic checks, resampling, model-as-judge, human review) on clinical adverse-event extraction from the CADEC corpus revealed they sorted answers by trustworthiness but corrected almost none, disproving the additive-accuracy theory. The zero-shot extractor achieved only ~33% accuracy, while deterministic checks caught nonexistent codes but could not validate plausible answers, and recognition tasks (recalling SNOMED codes from memory) invented 13-18% of codes versus menu-picking which scored ten times higher. Yield—the proportion of records with a correct answer—exposed what accuracy hides, and detection set a hard ceiling: no layer recovers a mention the extractor missed, so the largest gains came from matching annotation conventions rather than domain-adapted models.