AICoder

AI Coding Live News

First-hand AI coding updates from 𝕏/Twitter frontier labs and researchers, distilled by AI with entity benchmarks.

Explore live on 𝕏 ↗

AI Coding Live News

y
younes@yb2698·
👀25
🔥87
View on 𝕏 ↗

Hugging Face Ships Harness Adapters: RL-Trained LoRAs Inside Unmodified pi / OpenCode / Hermes, with safeagents Blocking Reward Hacks

On 2026-11-06 Hugging Face Harness Hub released Harness Adapters: one rank-32 LoRA per base model and coding harness (pi, OpenCode, Hermes), trained with GRPO on verifiable tasks inside unmodified harnesses. Vendor smoke evals claim fewer tool calls with equal or better accuracy. A safeagents proxy rule-checks every tool call and blocks reward hacks. Bases include LiquidAI LFM2.5-24B-A2B, LFM3-1.7B, and experimental TwIL-8B.

Hugging Face Ships Harness Adapters: RL-Trained LoRAs Inside Unmodified pi / OpenCode / Hermes, with safeagents Blocking Reward Hacks

Key Takeaways

  • Mechanics: one LoRA per model x harness (rank 32, all-linear), GRPO-trained inside unmodified pi / OpenCode / Hermes
  • Safety: safeagents proxy rule-checks tool calls (undeclared tools, test edits, reward-file writes, answer-source fetches); blocked rollouts score 0
  • Vendor smoke: fewer tool calls than the base with equal or better accuracy; one vLLM serves named adapters
  • Envs: SmolDataEnvs plus 230 Repo2RLEnvs tasks (repair/build/extend) rebuilt from public Hub trajectories
  • Open: blog, safeagents, training recipe, and live dashboard; follow runs at https://harness-hub-harness-adapters-rl.hf.space/
Read full story
Linked Models & Agents:
⚡ LFM2.5-24B-A2B⚡ LFM3-1.7B⚡ TwIL-8B🤖 pi🤖 OpenCode🤖 Hermes🌐 Hugging Face🌐 Liquid AI🌐 webAI
#Hugging Face#Harness Adapters#LoRA#RL#Safeagents#Pi#OpenCode#Hermes#agent tooling#GRPO
S
SlowMist@SlowMist_Team·
👀72
🔥98
View on 𝕏 ↗

LiteLLM Gateway Compromised in v1.107.0 Supply-Chain Attack: Emergency Rollback to 1.106.2 and Credential Rotation Issued

LiteLLM, a widely deployed unified proxy gateway powering millions of LLM and autonomous agent pipelines worldwide (including LangChain and OpenHands), suffered a critical supply-chain compromise. Attackers breached its automated CI/CD release pipeline, pushing a poisoned v1.107.0 release to PyPI and Docker Hub tags ('latest' and 'main-stable') equipped with a credential-harvesting payload targeting system environment variables and cloud API keys. Coordinated with security firms including SlowMist, the maintainers quickly contained the incident, yanked v1.107.0, restored PyPI to 1.106.2, and urged all operators to pin dependency versions and rotate API credentials immediately.

LiteLLM Gateway Compromised in v1.107.0 Supply-Chain Attack: Emergency Rollback to 1.106.2 and Credential Rotation Issued

Key Takeaways

  • LiteLLM release pipeline compromised, publishing trojanized v1.107.0 packages across PyPI and Docker Hub tags
  • Payload secretly exfiltrated host environment variables, harvesting model API keys, database credentials, and cloud secrets
  • Maintainers yanked the malicious package within hours, rolling PyPI and official Docker containers back to stable 1.106.2
  • Urges all agent and LLM platform operators to verify pinned package versions and execute immediate credential rotation
Read full story
Linked Models & Agents:
🤖 LiteLLM🤖 OpenHands🤖 LangChain🌐 BerriAI🌐 SlowMist
#LiteLLM#Security#Supply Chain Attack#Cybersecurity#API Key#DevOps
A
AikidoAikidoSecurity·
👀59
🔥88
View on 𝕏 ↗

LiteLLM 1.107.0 hit by a supply-chain compromise: PyPI wheels and Docker images shipped a credential stealer and backdoor; LiteLLM says it is contained and advises pinning 1.106.2

On October 21 the release pipeline of the popular LLM gateway LiteLLM was compromised. A malicious 1.107.0 was live on PyPI for about two hours (11:39 to 13:23 UTC), and the Docker tags v1.107.0, latest, main-latest and main-stable pointed to a poisoned image from 11:46 to 14:20 UTC. LiteLLM's incident report says the payload steals cloud, SSH, AI-tool and proxy keys on every Python start; the artifacts are pulled, OSS users should pin 1.106.2 and rotate secrets, and the private enterprise registry was not affected.

LiteLLM 1.107.0 hit by a supply-chain compromise: PyPI wheels and Docker images shipped a credential stealer and backdoor; LiteLLM says it is contained and advises pinning 1.106.2

Key Takeaways

  • Scope (vendor figures): all 7 PyPI wheels of litellm==1.107.0, live 11:39:51 to 13:23 UTC; 4 Docker Hub and GHCR tags poisoned from 11:46:56 UTC
  • Attack chain: the attacker held 3 sets of developer GitHub CLI credentials, one belonging to an org owner, used it to disable branch protection and force-push, then had the release workflow build and sign from the lookalike repo Berri-AI/litellm
  • Payload: cook.pth, cook.py and sitecustomize.py run on any Python start and survive uninstalling litellm; a second loader fetches remote code hourly
  • Fix: delete the 3 files and matching .pyc first, then pin litellm==1.106.2 or rebuild from scratch, and rotate every provider key in the proxy DB, the master key and the database password
  • This is LiteLLM's second supply-chain incident of 2026 (versions 1.82.7/1.82.8 in March); the GitHub source tag v1.107.0 itself is clean
Read full story
Linked Models & Agents:
🤖 LiteLLM Proxy🤖 Claude Code🌐 BerriAI🌐 PyPI🌐 Docker🌐 GitHub
#LiteLLM#supply chain attack#PyPI#Security#LLM gateway
S
SkyworkSkywork_ai·
👀105
🔥88
View on 𝕏 ↗

Skywork open-sources Matrix-Game 4.0: game engine owns state, a 5B causal video model renders playable games in real time on one GPU

Skywork AI (Kunlun Tech) open-sourced Matrix-Game 4.0 on 2026-10-15. An executable game engine owns world state, physics and rules and emits a cheap low-poly RGB proxy; a 5B causal video model built on Wan2.2-Fun-5B-Control re-renders it into finished frames with 1–2 denoising steps. Inference code, a WebRTC browser demo and game source (Apache-2.0) are on GitHub, with Castle, Ancient Egypt and City Racing weights on Hugging Face. Skywork claims 480p/60 FPS and 112 ms end-to-end on an RTX 5090; the repo's own A100 measurements are ~38–39 FPS. Training code and data are not released.

Key Takeaways

  • Scale: each scene checkpoint has 5,001,016,512 BF16 params on an alibaba-pai/Wan2.2-Fun-5B-Control base, output 480×864 (HF card)
  • Vendor claim (RTX 5090): 480p at 60 FPS, 112 ms end-to-end, up to 83.4 FPS generation; repo-measured on one A100-80GB: ~38–39 FPS for Castle/Egypt fixed prompts, ~28.9 FPS for Racing dynamic prompts (v4 README)
  • Presets: realtime = TAEHV codec + 1-step denoise + CUDA graphs; quality = Wan VAE + 2 steps; 21-latent attention cache incl. sink, 4 frames per chunk
  • License: code and scene weights Apache-2.0; no training code or data; weights need the Matrix-Game 4.0 runtime (not drop-in Diffusers)
  • Design: a harness where the engine owns state and rules, a coding agent turns natural-language requests into new rules and prompts, and the video model only renders
Read full story
Linked Models & Agents:
⚡ matrix-game-4.0⚡ wan2.2-fun-5b-control🌐 skywork🌐 kunlun-tech🌐 huggingface
#Skywork#Matrix-Game 4.0#Coding Agent#Hugging Face
ADSponsored
B
BerriAI@BerriAI·
👀97
🔥91

LiteLLM 1.103.2: forward Claude Code safeguards; attribute CLI session spend; restore pass-through pre-config wins

BerriAI shipped LiteLLM v1.103.2 on 2026-10-01 (PyPI ~06:20Z; 17 commits vs 1.103.1): forward Claude Code safeguards / dangerous-tool-use beta unchanged to Anthropic, Bedrock, Vertex, and Azure AI Foundry so auto mode works behind the proxy; attribute CLI session spend to a per-user cli-session alias for daily activity BI; restore pre-config-wins for pass-through endpoints.

LiteLLM 1.103.2: forward Claude Code safeguards; attribute CLI session spend; restore pass-through pre-config wins

Key Takeaways

  • Release v1.103.2 at 2026-10-01T06:37:31Z; 17 commits vs v1.103.1
  • Forward Claude Code safeguards + dangerous-tool-use beta to Anthropic/Bedrock/Vertex/Azure Foundry so auto mode returns safeguard_results
  • CLI session spend attributed to per-user cli-session alias for daily activity BI (#40541/#43642/#43656)
  • Restore pass-through pre-config-wins (#43962); verify Docker with cosign
  • pip install 'litellm==1.103.2'; https://docs.litellm.ai/docs/ ; https://pypi.org/project/litellm/1.103.2/
Read full story
Linked Models & Agents:
⚡ Claude Sonnet 5.5🤖 Claude Code (Anthropic)🌐 Anthropic🌐 Google DeepMind🌐 OpenAI
#LiteLLM#Proxy#Claude Code#Safeguards#Spend#Azure AI Foundry#Bedrock#Vertex#Release
O
OpenAI@openai·
👀94
🔥94

OpenAI Codex 0.159.0: opt-in instant_interrupt steering; warning dismiss; richer Mermaid & thread pagination

OpenAI Codex shipped rust-v0.159.0 on 2026-09-29 (npm @openai/[email protected]): opt-in instant_interrupt steers mid-response or long code-mode calls; compact welcome/headers; warnings viewer dismisses reviewed items on close (k to keep); scroll transcript while reviewing a plan; richer native Mermaid; app-server can paginate thread history from an item anchor. ~89 commits / 300 files vs v0.158.0. Windows suppresses stray consoles for MCP/code-mode; .aws protected under writable roots; macOS sandbox TLS/proxy fixes.

OpenAI Codex 0.159.0: opt-in instant_interrupt steering; warning dismiss; richer Mermaid & thread pagination

Key Takeaways

  • Release rust-v0.159.0 on 2026-09-29; ~89 commits / 300 files vs v0.158.0
  • Opt-in instant_interrupt steers mid-response / long code-mode (#48135/#48141)
  • Warnings dismiss on close (k to keep); scroll while reviewing plans; richer Mermaid
  • App-server paginates thread history from an item anchor (#48151)
  • Install: npm i -g @openai/[email protected]; removed tui.prompt_suggestions and bundled plugin-creator skill
Read full story
Linked Models & Agents:
🤖 Codex🤖 code-mode🌐 OpenAI
#OpenAI#Codex#CLI#MCP#Instant interrupt#Release
C
Cline@cline·
👀48
🔥90

Cline SDK 0.0.87: Windows proxy-safe hub startup; reasoning tokens not double-counted; Bedrock GPT-6 via inference profiles

Cline shipped SDK v0.0.87 on 2026-09-29: Windows hub startup works when HTTP(S)_PROXY is set—ensureLoopbackProxyBypass() adds loopback to NO_PROXY so Bun fetch no longer routes health probes through Clash/v2ray/corp proxies; normalizeUsage() excludes reasoning from outputTokens to stop double-counting; Gateway models keep per-model apiProtocol; bare GPT-6/GPT-5.6 on Bedrock use inference profiles (@ai-sdk/amazon-bedrock ^5.0.96); catalog 209→211 providers, 6,386→6,447 models; recommended list adds Claude Sonnet/Opus 5.5.

Cline SDK 0.0.87: Windows proxy-safe hub startup; reasoning tokens not double-counted; Bedrock GPT-6 via inference profiles

Key Takeaways

  • Release sdk/v0.0.87 vs v0.0.86
  • Windows proxy: ensureLoopbackProxyBypass() stops hub probes from going through HTTP(S)_PROXY
  • outputTokens excludes reasoning subset (cost still on full output); no double-count in session totals
  • Bedrock GPT-6/GPT-5.6 via inference profiles; @ai-sdk/amazon-bedrock ^5.0.96
  • Catalog 209→211 providers, 6,386→6,447 models; Sonnet/Opus 5.5 on recommended list
Read full story
Linked Models & Agents:
⚡ GPT-6⚡ GPT-5.6⚡ Claude Sonnet 5.5⚡ Claude Opus 5.5🤖 Cline🤖 Cline Hub🌐 Cline🌐 AWS Bedrock
#Cline#SDK#Bedrock#Proxy#Token usage#Release
B
BerriAI@BerriAI·
👀77
🔥94

LiteLLM v1.103.0 Released: Hosted vLLM Batches, MCP 2.x Architecture, and Admin UI Web Search Interception

LiteLLM shipped v1.103.0, introducing native hosted_vllm offline batch processing and updating its MCP integration test suite to the modern MCP 2.x MCPServer architecture. The Admin UI now supports dynamic web search interception controls, while the proxy resolves reasoning-object to reasoning_effort translation for OpenAI o1/o3 series and eliminates Azure 400 errors on empty tool choices.

LiteLLM v1.103.0 Released: Hosted vLLM Batches, MCP 2.x Architecture, and Admin UI Web Search Interception

Key Takeaways

  • Hosted vLLM Batching: Directly dispatches and tracks asynchronous batch inference across hosted vLLM clusters from unified LiteLLM client endpoints.
  • MCP 2.x Alignment & Budget Guardrails: Migrates test harnesses to the modern MCP 2.x MCPServer API while hardening zero-budget project request blocking and team-linked JWT inheritance.
  • Reasoning Effort & Azure Tool Fixes: Translates provider reasoning tokens into uniform reasoning_effort parameters and eliminates 400 errors on Azure tool_choice when no tools are attached.
Read full story
Linked Models & Agents:
#LiteLLM#vLLM#MCP#AI Gateway#Batch Processing
B
BerriAI@BerriAI·
👀83
🔥92

LiteLLM 1.103.0: MCP delegated OAuth admission, Fuse/Capability routers, prompt-cache cost prediction

BerriAI shipped LiteLLM v1.103.0 on 2026-09-28: delegated MCP OAuth now requires admission; Capability classifier + Fuse V2 routers land; the proxy can predict prompt-cache costs across deployments, bind JWT claims to agents, add tpd_limit, bulk user/team APIs, and /nvidia_nim passthrough. Also: Bedrock S3 managed file delete/list, Friendli price auto-sync, Responses↔Chat reasoning mapping, and cosign image verification docs.

LiteLLM 1.103.0: MCP delegated OAuth admission, Fuse/Capability routers, prompt-cache cost prediction

Key Takeaways

  • Shipped v1.103.0 — pip install litellm==1.103.0; docs.litellm.ai
  • MCP delegated OAuth requires admission; cosign-verify Docker images
  • Capability + Fuse V2 routers; cross-deployment prompt-cache cost prediction
  • JWT→agent binding, tpd_limit, bulk user/team management APIs
  • Bedrock S3 file delete/list, Friendli price sync, Codex model picker sync from proxy
Read full story
Linked Models & Agents:
🤖 Codex🌐 BerriAI🌐 AWS🌐 Friendli🌐 NVIDIA
#LiteLLM#MCP#OAuth#Router#Fuse#Prompt cache#Bedrock#Proxy#JWT
u
unclecode@unclecode·
👀74
🔥93

Crawl4AI 0.9.4: three Docker SSRF/env-leak advisories closed; PruningContentFilterLXML ~10× faster

unclecode shipped Crawl4AI v0.9.4 on 2026-09-23 (PyPI crawl4ai 0.9.4). It closes three coordinated advisories: blind SSRF via robots.txt (GHSA-f77g-77vp-r96v), SSRF with response disclosure via link_preview (GHSA-wh5w-hmj3-vgg7, high), and a dict-wrapper trust-boundary bypass that let non-admin clients read server env vars (GHSA-5w5p-vcv6-mm3f, high). A process-wide egress_policy wires the Docker PinningProxy into library HTTP clients. Default pruning switches to lxml-native PruningContentFilterLXML (~10×: 134→13 ms medium page; 2200→260 ms on a 6000-card page). Self-hosted Docker servers should upgrade.

Crawl4AI 0.9.4: three Docker SSRF/env-leak advisories closed; PruningContentFilterLXML ~10× faster

Key Takeaways

  • Shipped: v0.9.4 / PyPI 0.9.4; Docker: unclecode/crawl4ai:0.9.4
  • Security: three GHSAs closed; robots + link_preview SSRF and env-leak via dict laundering; pinning egress proxy
  • Perf: PruningContentFilterLXML default ~10× faster pruning with byte-identical output
  • Ops: CRAWL4AI_MAX_TIMEOUT_MS; pool max_pages_before_recycle=200
  • Install: docs.crawl4ai.com + pip install crawl4ai==0.9.4
Read full story
Linked Models & Agents:
🤖 browser-agent🤖 RAG-pipeline🌐 crawl4ai🌐 unclecode
#Crawl4AI#SSRF#Security#Docker#Web crawler#LLM#Pruning#Egress
ADSponsored
P
Prefect / FastMCP@PrefectHQ·
👀79
🔥89

FastMCP 4.0.10: task tools stay registered behind Search/CodeMode; cross-tool call_tool returns real results

Prefect shipped FastMCP v4.0.10 (2026-09-25, PyPI fastmcp 4.0.10, “Inside Job”): task-enabled tools remain registered with the task backend even when hidden by Search transforms or CodeMode, and calls via ctx.fastmcp.call_tool() (including the search call_tool proxy and CodeMode execute) now return real results instead of empty task receipts. Also tolerates stdio transport construction failures in __del__. Docs at gofastmcp.com.

FastMCP 4.0.10: task tools stay registered behind Search/CodeMode; cross-tool call_tool returns real results

Key Takeaways

  • Shipped: FastMCP v4.0.10 / PyPI fastmcp 4.0.10
  • Tasks: tools hidden by Search/CodeMode still register with the task backend (#5262)
  • Calls: ctx.fastmcp.call_tool() / CodeMode execute return real results, not empty task receipts (#5275)
  • Stdio: __del__ tolerates failed transport construction (#5256)
  • Docs: gofastmcp.com + repo changelog updated for v4.0.10
Read full story
Linked Models & Agents:
🌐 prefect
#FastMCP#MCP#CodeMode#Search transforms#task tools#Prefect#Python
C
Cline@cline·
👀62
🔥90

Cline Desktop v0.0.36: proxy-safe local hub, fit-to-screen UI, Bedrock GPT-6 without cross-region

Cline Desktop v0.0.36 (2026-09-25) skips system HTTP(S) proxies for local hub checks, fits oversized windows to small/scaled displays, loads plugin slash commands on workspace open, and runs Bedrock GPT-6/GPT-5.6 without cross-region inference (India via in. profiles).

Cline Desktop v0.0.36: proxy-safe local hub, fit-to-screen UI, Bedrock GPT-6 without cross-region

Key Takeaways

  • Local hub checks bypass system HTTP(S) proxies (Clash/v2ray/corporate)
  • Oversized windows shrink to fit and center on scaled 1080p displays
  • Plugin slash commands load on workspace open with background retry
  • Bedrock GPT-6/GPT-5.6 work without cross-region; India uses in. profiles
  • Artifacts: deb/rpm/dmg/exe + signatures; Providers panel scrolls model lists
Read full story
Linked Models & Agents:
⚡ GPT-6⚡ GPT-5.6🤖 Cline Desktop🌐 Cline🌐 Amazon Bedrock🌐 OpenAI
#Cline#Desktop#Proxy#Bedrock#GPT-6#Linux#coding agent
D
Dan Kornas@DanKornas·
👀53
🔥93
View on 𝕏 ↗

Clampdown open-sources a hardened container sandbox for Claude Code, Codex, and more

@DanKornas highlighted Clampdown (github.com/89luca89/clampdown), a hardened container sandbox for coding agents: Landlock filesystem rules before the binary starts, deny-by-default egress with an allowlist, real API keys held in a separate auth proxy while the agent gets a dummy key, OCI hooks for nested containers, plus launch helpers for Claude Code, Codex, OpenCode, and pi (GPL-3.0-only).

Key Takeaways

  • Landlock filesystem rules bind the agent to the project directory before start.
  • Real keys stay in an auth proxy; the agent gets a dummy key and deny-by-default egress.
  • Launch helpers for Claude Code, Codex, OpenCode, and pi; GPL-3.0-only.
Read full story
Linked Models & Agents:
🤖 Claude Code (Anthropic)🤖 codex
#Clampdown#Sandbox#Container#Landlock#Claude Code#Codex#Security#open source
C
Claude Code Log@ClaudeCodeLog·
👀50
🔥86
View on 𝕏 ↗

Claude Code 2.1.276 fixes 400 Input tag errors when ANTHROPIC_BASE_URL points at a proxy

Claude Code 2.1.276 is out about six hours after 2.1.275. The single CLI fix restores requests that failed with 400 Input tag "advisor_20260301" when ANTHROPIC_BASE_URL pointed at a proxy or gateway — a 2.1.275 regression.

Key Takeaways

  • Fixes the 2.1.275 regression where every request failed with 400 Input tag advisor_20260301 when ANTHROPIC_BASE_URL pointed at a proxy or gateway.
  • Single CLI change; critical for Claude Code users behind custom base URLs or enterprise proxies.
  • Changelog: anthropics/claude-code CHANGELOG.md#21276.
Read full story
Linked Models & Agents:
🤖 Claude Code (Anthropic)🌐 Anthropic
#Claude Code#2.1.276#CLI#ANTHROPIC_BASE_URL#Proxy#regression fix
J
Julius@jullerino·
👀49
🔥88
View on 𝕏 ↗

T3 Chat/Code: Limits under Usage with CLI Proxy hub and banked Codex resets

Julius (@jullerino) announced a new Limits section under Usage in T3 Chat / T3 Code: add a CLI Proxy hub for multi-account setups, and view/redeem banked Codex usage resets inside the app. Aimed at heavy Codex and CLI-proxy users managing quotas.

Key Takeaways

  • New Limits section under Usage centralizes quota settings.
  • Supports adding a CLI Proxy hub for multi-account/proxy setups.
  • View and redeem banked Codex resets inside the app.
Read full story
Linked Models & Agents:
🤖 openai-codex
#T3 Chat#T3 Code#Codex#CLI Proxy#Limits#Usage