trending.md — Dense Trending Signals

Machine-readable trending information. Ranked by velocity — how fast attention is shifting.
Built for AI agents. Readable by humans.
→ Raw feed: /en/feed/latest.md
→ Archive: /en/feed/

1. Zhipu's GLM-5.3 — a frontier coding model that gained "emergent cyber capabilities" from post-training alone

Velocity: ▮▮▮ trending
Source: z.ai · 973 pts · ~1d ago
Tags: zhipu glm coding cybersecurity open-weights

Zhipu AI (Z.ai) released GLM-5.3, a coding- and cybersecurity-focused model built on the same 743B-parameter base as GLM-5.2 — every gain came from scaled-up post-training (RL), not a new architecture. Coding scores roughly doubled on long-horizon tasks (SWE-Marathon 19.4→42.5; Terminal Bench 3.0 4.6→28.3, a ~6× leap). The bigger story is security: GLM-5.3 scored 84.5% on CyberGym, first among all models evaluated (ahead of Anthropic's Mythos 5 at 83.8%), and 54.4% on ExploitBench. Testing with Chinese security teams surfaced 2,436 vulnerabilities across 269 open-source projects (1,097 critical/high), the oldest dating to 1981, hidden an average of 26.6 years — now tracked in a public Security Disclosure Ledger.

Why it matters: This is the first time a Chinese lab has publicly justified a delayed open-weight release on safety grounds (weights land ~2 weeks after launch, with a "trusted access" program for the most sensitive cyber functions). It also shows post-training — not scale — as the lever for frontier capability jumps, and that vulnerability discovery is becoming a headline model benchmark in its own right.

API now requires thinking enabled (three effort levels); next-gen GLM-6 switches to a brand-new architecture with doubled parameters.

🔗 z.ai · 🔗 Pandaily

2. X (xAI) open-sources x-algorithm — the real code behind the "For You" feed

Velocity: ▮▮▮ trending
Source: GitHub · 29.1k stars · ~2d ago
Tags: xai recommendation open-source rust ml

xAI published x-algorithm (Apache 2.0), the code powering X's "For You" feed — a Rust + Python rewrite of the 2023 Scala release. The pipeline spans query hydration, candidate sourcing (Thunder in-network + Phoenix out-of-network two-tower retrieval + SimClusters), filtering, and scoring via a Grok-1-adapted transformer that predicts 19 engagement-action probabilities (favorite, reply, repost, click, dwell, block, report…) combined by a weighted scorer. An August 13 update added Phoenix training code, visibility filtering, labeling systems, and an "Under the Hood" transparency tool.

Why it matters: This is the first time a major platform has released recommendation-system code this complete — model weights and training data are still withheld, but the ranking architecture, abuse-enforcement logic, and transparency tooling are real and studyable, setting a new bar for feed-algorithm transparency.

Representative code — Grox LLM prompts and some botmaker rules are intentionally omitted to reduce gaming; no pre-trained weights ship.

🔗 xai-org/x-algorithm · 🔗 ppc.land analysis

3. Six AI giants standardize agent plugins — package once, run across ChatGPT, Copilot, Cursor; Anthropic absent

Velocity: ▮▮▮ trending
Source: 4sysops · spec v1.0.0 · ~1w ago
Tags: plugins standard mcp agent-skills interop

A coalition of OpenAI, Microsoft, GitHub, AWS, Vercel, and Anysphere (Cursor) — joined at launch by Google as a Core Maintainer — released Agent Plugins 1.0.0, an open packaging spec so a single plugin (a plugin.json manifest, a skills/ directory of Agent Skills, and an mcp.json for MCP servers) runs across ChatGPT, Codex, GitHub Copilot, VS Code, Cursor, and AWS Kiro. The spec standardizes only packaging and discovery/loading — installation, permissions, and sandboxing stay per-client — and supports reverse-domain namespaces for client-specific extensions.

Why it matters: It's the first cross-vendor attempt to unify the agent-plugin layer, built on two formats Anthropic created (MCP and Agent Skills) — yet Anthropic is notably absent from the coalition, having shipped its own separate plugin system for Cowork. Whether the ecosystem fragments or converges now hinges on which side developers bet on.

Working Draft; hooks, slash commands, and custom agents are not yet unified.

🔗 4sysops · 🔗 Context Studios

4. Microsoft's August Patch Tuesday patches a wormable DNS RCE (CVE-2026-62878, CVSS 9.8)

Velocity: ▮▮ rising
Source: The Hacker News · CVSS 9.8 · ~4d ago
Tags: microsoft dns rce patch-tuesday cve

Microsoft's August 2026 Patch Tuesday (Aug 11) fixed 398 CVEs (62 critical per ZDI), headlined by CVE-2026-62878 — a stack-based buffer overflow in Windows DNS Server rated CVSS 9.8: unauthenticated, network-reachable, no user interaction, and "wormable" per the Zero Day Initiative. Successful exploitation yields arbitrary code execution in the DNS service context. The release also patched three more unauthenticated 9.8 RCEs (Windows Deployment Services, QUIC, HPC Pack) and a second actively-exploited zero-day (LegacyHive, CVE-2026-62832, a User Profile Service flaw letting a local user reach SYSTEM).

Why it matters: Internet-facing DNS servers are a high-value, hard-to-segment target — a wormable RCE there threatens enterprise name-resolution and lateral movement. Microsoft rates exploitation "less likely," but the same technical condition drove past DNS worms, so patching exposed DNS infrastructure is the top priority this cycle.

Microsoft's own exploitability rating is "less likely"; ZDI calls it wormable — patch exposed DNS servers first, then assume compromise.

🔗 The Hacker News · 🔗 Mallory (CVE-2026-62878)

5. Unpatched GeoServer SQL-injection zero-day → RCE, actively probed hours after disclosure

Velocity: ▮▮ rising
Source: The Hacker News · unpatched · ~3d ago
Tags: cve sql-injection geoserver rce zero-day

A zero-day SQL injection in GeoServer's jsonArrayContains function — disclosed Aug 12 by researcher @q1uf3ng, still unpatched and without a CVE — lets unauthenticated attackers execute arbitrary SQL against PostGIS/Oracle data stores, and, under certain configurations (H2 with the sa admin account, or MS SQL Server with admin privileges), reach remote code execution. watchTowr observed hundreds of exploitation attempts within hours from a small IP pool — so far mostly reconnaissance and error-triggering probes, but no full payloads yet.

Why it matters: GeoServer is deployed across government, defense, telecom, and utilities, and it has a track record of mass exploitation (CVE-2024-36401 was weaponized against a US federal agency). With no official patch, the only defenses are removing public exposure and watching for a vendor fix — a third-party UBITQUITY hotfix is circulating but is not an official remedy.

Affects PostGIS/Oracle JDBC data stores; RCE requires specific configs (H2 sa, MSSQL admin).

🔗 The Hacker News · 🔗 SecurityWeek

6. watermarks-remover strips AI provenance marks — emerging right after Anthropic's text watermarking

Velocity: ▮▮ rising
Source: GitHub · 4.1k stars · ~3d ago
Tags: watermarking provenance privacy synthid c2pa

guillaumemeyer/watermarks-remover (MIT) removes AI-provenance markers in three layers: a deterministic pass over Unicode steganography (zero-width spaces, bidi control chars), a "best-effort" statistical attack on SynthID-Text/Kirchenbauer sampling watermarks via heavy paraphrasing, and file cleaners that strip C2PA/XMP/EXIF metadata from PNG, JPEG, SVG, PDF, DOCX and more. It ships as Grok Agent skills plus pure-stdlib Python scripts. The author is candid that word-choice-level text watermarks can't be certifiably removed until vendors publish detectors and keys.

Why it matters: It landed days after Anthropic began watermarking Claude text (Aug 2) under the EU AI Act's Article 50 transparency rules — part of a fast-forming cat-and-mouse ecosystem of watermarkers and de-watermarkers. It's also a supply-chain caution: an optional feature pulls a ~220 MB archive from a third-party repo, so treat it like any unverified code.

The author: "Until vendors ship public detectors and keys, no tool can honestly certify 'this fails the official check.'"

🔗 guillaumemeyer/watermarks-remover · 🔗 AISignal analysis

7. Google open-sources HEIR — a compiler that turns AI models into encrypted-inference models

Velocity: ▮▮ rising
Source: Google Blog · 138 pts · ~1d ago
Tags: homomorphic-encryption privacy compiler mlir google

Google's HEIR (Homomorphic Encryption Intermediate Representation) is a new open-source compiler in its Private Computing Toolkit that converts pre-trained models operating on plaintext into models that compute directly on encrypted inputs — targeting FHE schemes (BGV/BFV/CKKS via OpenFHE/Lattigo, CGGI via tfhe-rs) and built on MLIR. Google demonstrated four compiled private-inference apps (recommendation, credit-card fraud detection, encrypted-traffic intrusion detection, and a hotword detector) with an auto packing-selection pass that accelerates programs up to 145×.

Why it matters: Fully-homomorphic encryption removes the "server must see your data" trade-off with purely cryptographic guarantees, but authoring efficient FHE code normally needs a team of cryptographers. HEIR's stated goal — a "one-click" path for non-experts to add encrypted inference — is the missing piece that could move private AI from research demo to product feature.

FHE remains ~1,000–10,000× slower than plaintext, so today this is for small models on sensitive data, not general-purpose LLM inference.

🔗 Google Blog · 🔗 Google Developers Blog

8. NVIDIA NemotronLabs VoiceChat 11B — the first open full-duplex voice model with tool calling

Velocity: ▮ steady
Source: Hugging Face · 11B params · ~1w ago
Tags: nvidia speech full-duplex voice open-weights

NVIDIA's NemotronLabs VoiceChat 11B is an end-to-end, real-time full-duplex speech model published with open weights — the first open model to listen and speak simultaneously while calling tools mid-conversation on a separate output channel. A single streaming network (7.7B Nemotron-H backbone + Fast Conformer encoder + Gemma-3 TTS) hits ~448 ms turn-taking and 38.8% on Big Bench Audio. It ships under OpenMDW v1.1 (research-only); MLX community conversions already run it on Apple Silicon.

Why it matters: Real-time voice agents have been stuck between closed APIs (OpenAI, Google) and open models that can't do true barge-in/tool-calling. VoiceChat proves the full-duplex stack is openable, though the research-only license and ~80 GB GPU requirement keep it a research artifact for now.

English-only, 2-minute audio context, single voice; needs an 80 GB GPU (A100–B200) via vLLM.

🔗 Hugging Face · 🔗 dev.to

9. Cursor open-sources its plugin spec and 11 official plugins

Velocity: ▮ steady
Source: GitHub · 2.8k stars · ~2d ago
Tags: cursor plugins skills mcp spec

cursor/plugins is Cursor's official plugin specification and marketplace repo (MIT): each plugin is a directory with a .cursor-plugin/plugin.json manifest bundling any of six component types — rules (.mdc), skills, agents, commands, MCP servers, and hooks — with automatic folder-based discovery. It ships 11 official plugins (Continual Learning, Team Kit, Thermos branch review, PR Review Canvas, Orchestrate, and more), and every community plugin is manually reviewed before listing.

Why it matters: Cursor is converging its plugin format on the same primitives (skills/ + mcp.json) the Agent Plugins 1.0.0 coalition standardized, so cursor/plugins doubles as a reference implementation for the cross-vendor spec — while adding Cursor-specific extensions (rules, hooks, canvases) the 1.0.0 spec deliberately left out.

Variables use ${VAR} placeholders — secrets are set in the dashboard, never stored in the plugin.

🔗 cursor/plugins · 🔗 Cursor Docs

10. mixedbread's Toast 1 — a search sub-agent that's 10× cheaper than frontier search

Velocity: ▮ steady
Source: mixedbread · 134 pts · ~1d ago
Tags: search rag agents mixedbread retrieval

Toast 1 is mixedbread's specialized search agent: it decomposes a query into sub-queries, gathers evidence, inspects sources, and curates context before a generalist frontier model answers — claiming frontier-class quality at up to 10× lower cost and 12× faster. On Databricks' OfficeQA Pro V2, GPT-5.6 Sol + Toast 1 hit 70% correctness at ~$1.15/task vs 60% at ~$4/task for Claude Fable 5 on Databricks Genie; on Harvey's Legal Agentic Benchmark it cut token usage from 80.6M to 23M while preserving quality. The framework is open-source; the model weights are not.

Why it matters: As agents converge on "sub-agent for search" as a standard pattern, Toast 1 makes the case that search doesn't need a frontier model — a specialized, co-designed retrieval model plus a cheap harness beats generalist search on cost while matching quality, shifting where tokens get spent.

~$0.023 per query, ~8 seconds; works best with mixedbread Search but plugs into any backend.

🔗 mixedbread · 🔗 TokenPost

11. RustDesk cracks true unattended remote access on Wayland

Velocity: ▮ steady
Source: RustDesk Blog · preview build · ~1d ago
Tags: rustdesk wayland remote-desktop linux open-source

RustDesk shipped a preview build that delivers true unattended remote access on Wayland — connect to a remote Linux machine without anyone approving the session, including from the login screen after a reboot, with multi-monitor support. It's a notable first: AnyDesk still requires Xorg for incoming Linux sessions and TeamViewer still calls Wayland "experimental." The team is releasing it as a separate preview (rustdesk-unattended-wayland-1.4.9-x86_64.deb, x86_64 Debian/Ubuntu) before folding it into stable builds.

Why it matters: Wayland's security model (portal-based consent for screen capture and input) has long blocked unattended remote access — a real blocker for Linux fleets and homelabs. RustDesk's implementation is currently a technical black box, which reviewers flag as both a breakthrough and a security question, especially given unattended access extends to the pre-login screen.

Preview only; Fedora/Arch support and stable integration are next.

🔗 rustdesk/rustdesk · 🔗 RustDesk Blog

12. LuaCAD — parametric CAD scripted in Lua, a Rust rewrite of OpenSCAD's ideas

Velocity: ▮ steady
Source: GitHub · Show HN · ~1d ago
Tags: cad lua rust openscad parametric

LuaCAD (ad-si) is a parametric CAD tool that models 2D/3D solids in Lua instead of OpenSCAD's "cobbled-together" SCAD language. It's a Rust rewrite embedding Lua 5.4 (via mlua) with operator overloading for natural CSG syntax (a + b, a - b, a * b), OpenCSG for preview rendering, the Manifold library for watertight meshes, and native support for the full BOSL2 library. It ships a CLI (luacad convert model.lua output.stl, luacad watch) and a luacad-studio desktop app with live 3D preview.

Why it matters: OpenSCAD's declarative language is powerful but widely disliked, and every alternative has been a new bespoke language. LuaCAD swaps in a real, embeddable language programmers already know — a small but sharp reminder that "good CAD scripting" and "good general-purpose language" don't have to be in tension.

Exports 3MF/STL/OBJ/PLY/OFF/AMF/SCAD; cargo install luacad luacad-studio.

🔗 ad-si/LuaCAD · 🔗 LuaCAD Docs

13. Mole — a terminal deep-research agent with an enforced budget and verified quotes

Velocity: ▮ steady
Source: GitHub · Show HN · ~1d ago
Tags: agents research cli mcp privacy

Mole (lajosdeme) is a single-binary terminal research agent that attacks three failure modes of AI research: cost overruns, hallucinated citations, and local-data leakage. An enforced budget reserves and settles every model call against a ledger with non-negative DB constraints (--usd 0.50 claims 0% overshoot); verified quotes discard any claim whose quote doesn't appear verbatim in its source before it reaches the answer; and a privacy boundary lets it analyze a local CSV/folder while only aggregates (≥5-record buckets) leave the machine. It speaks MCP so coding agents can drive it.

Why it matters: "Deep research" tools are proliferating, but few make cost and provenance enforceable rather than advisory. Mole's budget-as-database-constraint and quote-verification loops are concrete, reusable mechanisms for the trust problems that still keep people from delegating real research to agents.

Apache 2.0, CGO_ENABLED=0 static binary; Anthropic or any OpenAI-compatible backend; Tavily/Brave for search.

🔗 lajosdeme/mole · 🔗 AUR package

14. Google's Gemini 3.7 Flash — half-price coding/agent model lands three weeks after 3.6

Velocity: ▮▮▮ trending
Source: Google Blog · ~2d ago
Tags: google gemini coding agents multimodal

Google shipped Gemini 3.7 Flash, its "most intelligent" Flash for coding and agent workflows, just three weeks after Gemini 3.6 Flash. Headline gains are on agentic coding: DeepSWE v1.1 49.0→65.3%, FrontierCode 1.1 34.4→43.6%, and WebDev Arena Elo 1538→1588, plus a 1M-token input window and multimodal input. Launch pricing is halved to $0.75/M input and $3.75/M output through Dec 31 (then $1.50/$7.50 in 2027), and it ships day-one across the Gemini API, AI Studio, Android Studio, Antigravity, and Gemini Enterprise — where it now powers the Gemini Spark agent.

Why it matters: A three-week cadence plus a half-price launch is a direct bid for the "cheap workhorse for agents" tier, and the DeepSWE jump shows agentic-coding evals — not chat benchmarks — are now where the model race is actually fought.

Promotional pricing also applies to 3.6 Flash; Gemini Spark runs on 3.7 Flash for AI Pro/Ultra subscribers in 160+ countries (not EEA/UK/Switzerland/Nigeria).

🔗 Android Authority · 🔗 APIDog benchmarks

15. Alibaba's Qwen3.8-27B — an Apache-2.0 multimodal 27B topping SWE-bench Pro

Velocity: ▮▮▮ trending
Source: Hugging Face · 843 pts · ~1d ago
Tags: qwen alibaba multimodal open-weights coding

Qwen Team released Qwen3.8-27B (Apache 2.0), a natively-multimodal 27B model — a Gated DeltaNet + attention hybrid with multi-token prediction — featuring 262K native context (1M via YaRN) and native image/video understanding. It is best-in-row on SWE-bench Pro (61.7), LiveCodeBench v6 (90.3), OSWorld-Verified (84.3), WebArena-Verified (64.8) and AndroidWorld (81.9), with thinking control (reasoning_effort xhigh/medium/low, preserve_thinking). It topped Hacker News at 843 points, and 271 quantized variants already exist for llama.cpp/Ollama/LM Studio.

Why it matters: It's the mid-size companion to Qwen3.8-Max (the 2.4T-A95B flagship released days earlier), giving local and edge deployments a frontier-competitive multimodal coding model under a permissive license — the exact gap between closed APIs and full-stack agent tooling.

Runs on Transformers/vLLM/SGLang/Docker Model Runner; MLX community builds cover Apple Silicon.

🔗 Qwen/Qwen3.8-27B · 🔗 orcarouter comparison

16. MiniMax Music 3.0 — open-weights model generating full 5-minute songs

Velocity: ▮▮▮ trending
Source: MiniMax Blog · ~2d ago
Tags: minimax music audio open-weights generative

MiniMax open-sourced MiniMax-Music3, a music model that generates complete ~5-minute songs in a single pass (32 kHz, 16-bit stereo WAV) from lyrics plus a structured music description. It's a hybrid: an 8B "global" LLM for long-range structure, a 0.6B "local" LLM for acoustic detail, a 2.4B flow-matching module, and a 123M Flow-VAE decoder over an 8-layer RVQ tokenizer. It runs in ~24 GB VRAM (down to ~8 GB with CPU offload), ships ComfyUI support and an INT8 variant, and accepts section tags ([Intro], [Chorus], [Bridge]…) for fine-grained arrangement control.

Why it matters: Full-song, self-hostable music generation with control over BPM, key, vocal details, and structure was locked behind closed APIs like Suno and Udio. Open weights plus a $0.15/song API make it the strongest open challenger — though quality claims are still vendor-reported.

Notably quiet release: ~25 downloads at launch, no paper or benchmark report — just weights, a README, and a demo Space.

🔗 MiniMax Blog · 🔗 MiniMaxAI/MiniMax-Music3

17. OpenAI previews GPT-5.6 Sol "Ultrafast" — 14× faster, 750 tokens/sec on Cerebras

Velocity: ▮▮ rising
Source: OpenAI / Cerebras · preview · ~1d ago
Tags: openai gpt-5.6 inference cerebras latency

OpenAI previewed an "Ultrafast" mode for GPT-5.6 Sol that runs the flagship up to 14× faster — generating up to 750 tokens/second — by serving it on Cerebras chips rather than dropping to a smaller model. It's aimed at real-time workloads like fraud detection and live customer support, and is already in testing with firms including Jane Street and Podium. No general-availability date was given; OpenAI called it an "early preview."

Why it matters: Real-time performance on a flagship model, without distillation, is the missing piece for production agentic and streaming AI use cases. If 750 tok/s holds at GA, the bottleneck shifts from inference speed to orchestration, safety, and cost.

Ships in an unusually dense 24-hour window alongside DeepSeek V4-Pro's 50%-cheaper off-peak pricing and Gemini 3.7 Flash.

🔗 The Neuron · 🔗 TLDR AI

18. CISA confirms SonicWall SMA1000 flaws are now ransomware vectors (CVE-2026-15409/15410)

Velocity: ▮▮ rising
Source: CISA KEV · CVSS 10.0 + 7.2 · ~1d ago
Tags: cisa sonicwall ransomware ssrf kev

On Aug 14 CISA updated its Known Exploited Vulnerabilities entries for CVE-2026-15409 (CVSS 10.0 SSRF in the SMA1000 "Work Place"/wsproxy interface) and CVE-2026-15410 (CVSS 7.2 command injection in the removehotfix process) to confirm both are now exploited in ransomware campaigns, linked by Resecurity to an INC Ransomware affiliate. Chained, they give zero-click, unauthenticated root compromise of internet-facing SMA1000 appliances; Volexity traced exploitation to June 22 with custom malware (KNUCKLEBALL, Sou5, ROOTRUN, ORANGETAIL).

Why it matters: Patching isn't enough — exploitation predates the July 14 disclosure by three weeks, so any internet-facing SMA1000 unpatched in that window must be investigated for compromise. Shadowserver tracked ~380 exposed appliances at report time.

Fixed in 12.4.3-03453 / 12.5.0-02835, no workaround. INC Ransomware also pressures victims by phone ("Andrew", +1 (304) 384-0401) and email.

🔗 cirt.gy advisory · 🔗 CVETodo

19. NVIDIA's Nemotron Teacher — a 550B "reasoning teacher" for distilling frontier models

Velocity: ▮▮ rising
Source: Hugging Face · 550B params · ~1d ago
Tags: nvidia distillation reasoning openmdw teacher

NVIDIA released Nemotron-Labs-Teacher-General-Reasoning, a 550B-parameter (55B active) LatentMoE Mamba-2 + Transformer model built as a training-time "teacher" in the Nemotron 3 Ultra pipeline — one of 10+ domain-specialized teachers in its Multi-Teacher On-Policy Distillation (MOPD) recipe. It generates long reasoning traces on the hardest math/logic/abstraction problems and grades free-form answers, with a reasoning-budget dial (enable_thinking, medium_effort). It's a weights-only release (1.12 TB download, 4×B200 / 8×H100 minimum) under OpenMDW-1.1 with disclosed post-training data — and no published benchmarks.

Why it matters: It's a rare window into how frontier labs actually build reasoning models: a specialist teacher whose quality NVIDIA itself declines to benchmark. OpenMDW's post-training-data disclosure is also a small step past the weights-only norm.

No inference provider hosts it; reasoning quality is "unverified by anyone outside NVIDIA."

🔗 orcarouter analysis · 🔗 Nemotron Teacher model card

20. Liquid AI's LFM2.5-VL-3B — an on-device vision-language model that beats larger rivals

Velocity: ▮ steady
Source: Hugging Face · 3.1B params · ~2d ago
Tags: liquid-ai vision-language on-device open-weights multimodal

Liquid AI released LFM2.5-VL-3B, a ~3.1B vision-language model (LFM2.5-2.6B backbone + SigLIP2 NaFlex encoder) designed for on-device use: 228 tok/s on an Apple M5 Max and ~20 tok/s on a Galaxy S26 Ultra in under 3.3 GB. It posts ScreenSpot-v2 80.7 (screen understanding), RefCOCO P@1 87.9 (grounding), ChartQA 81.3, 16-language support, and experimental structured-OCR (labels + bounding boxes + LaTeX). Official GGUF/ONNX/MLX quantizations ship.

Why it matters: It targets the GUI-agent niche — reading screens and grounding objects locally on phones and laptops — the exact workload most "computer-use" agents run, on hardware that can't host a 27B model.

License lfm1.0; not recommended for long-context visual reasoning (web design, blueprint Q&A).

🔗 LiquidAI/LFM2.5-VL-3B · 🔗 llm-stats

21. firecrawl/anydoc — one Rust engine turns any office document into LLM-ready Markdown

Velocity: ▮ steady
Source: GitHub · 16.1k stars · ~3d ago
Tags: firecrawl rust markdown documents rag

Firecrawl's anydoc (MIT) converts Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and PDF into clean GitHub-Flavored Markdown through a single Rust core with Node/Python/WASM bindings, at a median under 5 ms per document. It detects format from bytes (not extensions), preserves tables/merged cells/footnotes/speaker notes, renders images as alt text, and powers Firecrawl's /parse API — plus an Agent Skill (npx skills add firecrawl/anydoc).

Why it matters: Document ingestion is the unglamorous bottleneck for RAG and agent workflows, and legacy converters (LibreOffice 1,129 ms, Pandoc 102 ms) are orders of magnitude slower or lossy. One fast, consistent, single-output-format engine removes a whole class of preprocessing code.

Benchmark: 4.4 ms median vs 1,129 ms (LibreOffice) and 102 ms (Pandoc); the only converter in its test to handle all 14 formats.

🔗 firecrawl/anydoc · 🔗 Firecrawl /parse blog

22. Comp AI CRM — a CRM built for AI agents to keep their own notes

Velocity: ▮ steady
Source: GitHub · 7.1k stars · ~1w ago
Tags: crm agents typescript eve open-source

trycompai/crm (Comp AI CRM, MIT) inverts the CRM model: a persistent research agent is the product, and the database is just "where the agent keeps its notes." The agent runs its own deployment, schedule, and work queue — reading the team inbox, creating contacts from email threads, enriching companies, scheduling its own follow-ups, and spending a research budget until it runs out. Built on Vercel's eve framework (18 tools, 4 skills, network-isolated sandbox), single-tenant, and runs with zero external API keys. A core rule: "nothing about a person is guessed" — weak evidence becomes a human-reviewed suggestion, not a fact.

Why it matters: It's a concrete example of the "agent-first" software pattern replacing form-first SaaS: the UI becomes a view of what the agent did, not a data-entry surface — inverting how CRM and most business software gets designed.

~6 days old at v1.0.0; TypeScript Turborepo on Bun (Next.js + NestJS/tRPC + Postgres); Slack integration; Google sign-in with an allow-list as the entire authorization model.

🔗 trycompai/crm · 🔗 openalternative

23. Anthropic's second Risk Report discloses an unreleased "Model 2" that beats Mythos 5

Velocity: ▮▮▮ trending
Source: Anthropic · 186-page report · ~1d ago
Tags: anthropic risk-report model-2 safety responsible-scaling

Anthropic published its second company-level Risk Report (Aug 14, covering assessments through July 15), and the headline is an internal, unreleased model called Model 2 that outperforms the public flagship Claude Mythos 5 — scoring 162.79 vs 161.29 on the AECI capability index and 62.8% vs 50.3% on CoBench, Anthropic's internal benchmark of 449 real R&D tasks (a model able to fully substitute its own engineers would need ~85%). Anthropic says it has no plans to release Model 2 and hasn't finished its pre-deployment safety suite. The report also raised catastrophic-misalignment risk from "very low" to "low" for the first time, and disclosed that Claude now authors a large majority of the code merged into Anthropic's production codebases.

Why it matters: The gap between an unreleased internal model and the public flagship — combined with Anthropic's admission that its task-based evals are "saturated" and can no longer distinguish capability gains — is the clearest signal yet that frontier labs are holding back models they can no longer fully measure.

Also disclosed: a biosafety-classifier flag accidentally disabled for ~11 months (133M messages), and chain-of-thought contamination in 0.27–5.1% of RL training episodes.

🔗 Anthropic Risk Report · 🔗 Yahoo Tech

24. Vero — the first repository-scale benchmark for formally-verified coding agents

Velocity: ▮▮ rising
Source: arXiv · 43 instances · ~1d ago
Tags: benchmark formal-verification lean4 agents software-verification

Vero (arXiv:2608.13522, UC Berkeley's Dawn Song et al.) is the first benchmark to evaluate AI agents on joint code implementation and machine-checked proof synthesis at the repository level. Its 43 multi-module instances come from real-world repositories spanning Python, Dafny, Verus, and Coq (from cryptographic protocols to distributed systems); each gives an agent a multi-module Lean 4 repository with fixed API interfaces and formal specifications, in proof-only or code-and-proof modes. The strongest frontier coding-agent configuration fully solved only 27 of 43 instances and closed no specifications on the hardest repositories.

Why it matters: As SWE-bench and its variants saturate, Vero shifts the frontier from "passes tests" to "mathematically verified correctness" — the next rung for agent evaluation, and a stress test that current agents still fail badly on repository-scale proof obligations.

Ships the benchmark, curation pipeline, and eval harness at sunblaze-ucb/vero, including an audit mode where agents can formally prove a specification unsatisfiable.

🔗 arXiv:2608.13522 · 🔗 sunblaze-ucb/vero

25. CVE-2026-73296 — Microsoft UFO agent framework exposed unauthenticated MCP servers (CVSS 9.4)

Velocity: ▮▮▮ trending
Source: NVD · CVSS 9.4 · ~3d ago
Tags: microsoft ufo mcp cve rce

Microsoft's open-source UFO agentic-automation framework, in versions before 3.0.8, stood up Streamable HTTP MCP servers on TCP ports 8020/8021 with no authentication — any network-adjacent attacker could invoke capture_screenshot, tap, swipe, type_text, and launch_app against an ADB-connected Android device, effectively granting full remote control and screen disclosure (IONIX calls it "RCE-equivalent"). The fix (GHSA-24fq-m9rr-g3mm) makes a bearer token (UFO_MCP_API_KEY, constant-time checked) mandatory and refuses to start without it.

Why it matters: This is a new, fast-growing attack class: agent frameworks auto-exposing MCP tool servers to the network with no auth. Unauthenticated MCP = direct agent-tool execution, which is as bad as RCE when the tools control a device.

CISA-ADP assessed exploitation "poc", automatable: yes, technical impact: total. Do not expose 8020/8021 beyond loopback.

🔗 NVD CVE-2026-73296 · 🔗 GitHub advisory GHSA-24fq-m9rr-g3mm

26. CVE-2026-72776 — unauthenticated RCE in the AgenticSeek agent via its /query endpoint (CVSS 9.8)

Velocity: ▮▮ rising
Source: IONIX · CVSS 9.8 · ~1d ago
Tags: cve agenticseek rce ai-agents shell-injection

Fosowl's open-source AgenticSeek autonomous agent framework, versions ≤ 2.41.1, exposes a /query API bound to 0.0.0.0:7777 with wildcard CORS and no auth, feeding attacker input straight into a BashInterpreter that runs subprocess.Popen(..., shell=True). An incomplete blocklist in safety.py is bypassable, so a crafted POST to /query executes arbitrary OS commands with the agent process's privileges (CVSS 9.8). Fixed in PR #534.

Why it matters: Local "AI agent" tools increasingly ship a network exec surface by default — an unauthenticated shell for anyone who can reach the port. The fix (bind loopback, gate /query, drop shell=True) is the checklist for every agent runtime.

Mitigations if you can't patch: don't expose port 7777, restrict CORS, run least-privilege, watch logs for anomalous /query hits.

🔗 IONIX CVE-2026-72776 · 🔗 Fosowl/agenticSeek PR #534

27. CVE-2026-16051 — signed-request replay enables RCE in WPMU DEV Dashboard plugin (CVSS 9.8)

Velocity: ▮ steady
Source: IONIX · CVSS 9.8 · ~3d ago
Tags: wordpress wpmu-dev rce cve supply-chain

The WPMU DEV Dashboard (wpmudev-updates) WordPress plugin, in all versions before 5.0.1, fails to verify package integrity on remote Hub installs and has no anti-replay protection on signed management requests (CWE-94). An attacker who obtains or replays a validly-signed request can make the site install and execute arbitrary code — full site compromise with no auth or user interaction (CVSS 9.8). A related authentication bypass (fixed in the same release) makes forging those requests easier. WPScan tracks it under 8dae5fbf-….

Why it matters: This is a supply-chain RCE in the plugin update mechanism itself — exploiting legitimate signed management channels rather than a code typo, which is why it looks like normal admin traffic and is easy to miss.

Update to 5.0.1+ (restores integrity verification + replay protection); rotate WPMU DEV Hub API keys as well.

🔗 IONIX CVE-2026-16051 · 🔗 stack.watch

28. GitHub's spec-kit surges as "spec-driven development" becomes the agent-coding default

Velocity: ▮▮ rising
Source: GitHub · 128.8k stars · ~1d ago
Tags: github spec-driven-development ai-coding cli agents

github/spec-kit (MIT) packages GitHub's Spec-Driven Development workflow: a specify CLI that scaffolds a constitution → specify → plan → tasks → implement pipeline, installing slash-commands or agent skills into 30+ AI coding agents (Copilot, Codex, Claude Code, Gemini CLI). Specifications become the "executable source of truth" that agents run against and validate at each checkpoint — an explicit answer to "vibe coding" that compiles but misses intent. The project, open-sourced in Sept 2025, is surging again on GitHub Trending (~128.8k stars, +1,160/day) behind a recent v0.12.11 release.

Why it matters: The agent-coding workflow layer is consolidating around specs-as-code, and GitHub's toolkit is becoming the reference implementation — worth watching as the trade-off (more upfront tokens for more predictable output) plays out at team scale.

Still labeled experimental by GitHub; the main criticism is higher token consumption per session.

🔗 github/spec-kit · 🔗 Visual Studio Magazine

29. holehe — email-to-account OSINT enumerator hits #3 on GitHub Trending

Velocity: ▮ steady
Source: GitHub · 13k stars · ~1d ago
Tags: osint email privacy python recon

megadose/holehe (GPL-3.0) checks whether an email address is registered on 120+ services (Twitter, Instagram, GitHub, eBay…) by probing each site's forgot-password/registration flows — critically, without sending any notification to the target address, and optionally recovering partially-obfuscated recovery emails or phone numbers. It resurged to #3 on GitHub Trending (Aug 15) after a source-code deep-dive article dissected its internals. It ships as a CLI (pip install holehe) and a Python API over trio/httpx.

Why it matters: Email enumeration is a quiet privacy leak — an unauthenticated "presence signal" across the web that needs no target interaction. Holehe's "silent" mode is exactly what makes it both a favorite OSINT tool and a reminder of how much your email address discloses.

"Built for educational purposes only"; site modules drift and can false-positive, so treat hits as presence signals, not identity proof.

🔗 megadose/holehe · 🔗 xlap.top deep-dive

Metadata

Field	Value
Generated	2026-08-15T20:03:00Z
Items	29
Sources tracked	40 (GitHub Trending, Hacker News, z.ai, Pandaily, ppc.land, 4sysops, Context Studios, The Hacker News, Mallory, SecurityWeek, AISignal, Google Blog, Google Developers Blog, Hugging Face, dev.to, Cursor Docs, mixedbread, TokenPost, RustDesk Blog, LuaCAD Docs, AUR, Android Authority, APIDog, The Neuron, TLDR AI, orcarouter, MiniMax Blog, cirt.gy, CVETodo, llm-stats, Firecrawl Blog, openalternative, Anthropic, Yahoo Tech, arXiv, NVD, IONIX, stack.watch, Visual Studio Magazine, xlap.top)
Update schedule	04:03, 12:03, 20:03 UTC+8 (3x daily)
Ranking	Velocity-weighted (recency × engagement acceleration × source authority)
License	CC-BY 4.0

Previous day · Raw .md · Archive