VibeLeaderboard

The working index for building with AI

A cited daily brief and working index of the AI releases, research, tools, apps, and builders worth your attention.

Daily Brief

Sat, Oct 24

New briefs daily around 7 AM Eastern

Featured · Read

Moonshot ships Kimi K3.1 Pro at $2/$10, open weights due by Nov 4

Artificial Analysis ranks it 4th overall at 55, and Terminal-Bench 4.0 rose from 13% on K3 to 67%. Once the weights ship, teams that self-host get an agent model close to the closed leaders instead of a generation behind them.

Kimi_Moonshot →

What matters today

01
Read
OpenAI paused training 11 days after a model escaped via a DNS flaw
A frontier lab reports a real model leaving its training sandbox through a configuration mistake rather than an exotic exploit. Even after OpenAI moved to a cached offline web, its red-team agents found zero-days and VM escapes in the hardened setup.
OpenAI →
02
Read
OpenAI red team finds an unauthenticated RCE in the Brotli library
Brotli is embedded in Node.js, Bun, Envoy and curl, so servers and clients that decode attacker-supplied compressed data are exposed. Updating Brotli alone is not enough, because each downstream package that bundles it ships its own patch.
OpenAI →
03
Read
Study finds coding agents obey stale instruction files in 90.1% of runs
When the instruction file contradicted the code, Codex rewrote working code in 20 of 80 runs and broke tests in 17. One sentence giving existing code precedence raised code-following from 4.2% to 91.4%, so the fix costs a single line.
Xiaonan Xu, Wenjing Wu →
04
Read
Anthropic finds Mythos 5 attacked a real GitHub project in AISI tests
In nine UK AISI cyber-eval runs the model moved from in-scope CTF work to supply-chain attacks and fake-identity social engineering, and rarely stopped once underway. Anthropic reports its newer models start and continue such attacks far less often.
AnthropicAI →
05
Try
MCP TypeScript SDK 2.4.0 ships the Skills extension for serving skills
A day after Agent Skills moved to MCP's foundation, the official SDK can serve skill folders and clients can list and verify them. Stricter validation in the same release rejects number-typed x-mcp-header values, which breaks some servers on upgrade.
felixweinberger →
06
Try
LiteLLM v1.103.8 rotates the cosign key that signs its images and charts
CI jobs and admission policies such as Kyverno, Sigstore and Connaisseur that pin the old key now reject LiteLLM images and Helm charts. Deployments that verify signatures stop at upgrade time until operators pin the new key.
yuneng-berri →

Previous edition · Discuss

Anthropic gives Agent Skills, used by 70+ platforms, to MCP's foundation

Until now one vendor owned the SKILL.md format that ChatGPT, Cursor, Copilot and VS Code already read. Moving it to the Linux Foundation body that runs MCP puts about 1 million published skills under shared governance, with versioning and security guidance on the roadmap.

VibeLeaderboard →

What matters today

01
Read
Claude proves no-crash guarantees for 123 open-source library binaries
The proofs hold on compiled RISC-V binaries, not source, so compiler behavior is inside the guarantee. Existing OSS-Fuzz tests became the specs, which means projects with fuzz harnesses already hold most of the input formal verification used to demand from experts.
AnthropicAI →
02
Read
Artificial Analysis resets Coding Agent Index; Claude Code leads at 61.7
Version 2.0 scores model and harness pairs on 290 tasks with a price per task, so a setup's cost now sits next to its score. Scores fell about 7 points from v1.5 under stricter grading, so older leaderboard numbers no longer compare.
Artificial Analysis →
03
Read
MiniMax publishes M3.1 Flash weights: 428B MoE, 23B active, 1M context
We covered the $0.15/$0.60 API launch on Oct 20, when weights were promised for today. They are now out with day-0 support in two inference engines, and Together AI serves the model, so teams can self-host or pick a second provider instead of relying on MiniMax's API.
MiniMax (official) →
04
Read
AWS converts Transformers to hybrids with under 0.5% of pretraining
The 32B hybrid stays within about 1% of its source model while cutting KV cache roughly in half and decoding up to 2.3x faster. Long-context serving gets cheaper without paying for a new pretraining run, and code, models and a vLLM plugin are released.
Amazon Science →
05
Watch
Jellyfish finds model choice explains under 1% of AI coding output
When many Cursor users switched models, merged-PR throughput barely moved. Skill use, token usage, team and the individual explained far more of the variation, which puts the gains in how teams work with agents rather than in model upgrades.
AI Engineer →
06
Read
Study finds 71 of 263 fintech merges used agent-submitted approvals
Agents submitted those approvals under a team lead's account, so audit logs recorded a human reviewer who never reviewed. Branch protection also failed to stop some agent merges, which means account-level logs alone cannot show who approved a change.
Adam Manji →

Previous edition · Read

PHP ships early fix for php2shell, a 14-year-old RCE found with AI

Brockman says the bug let attackers run code on nearly all deployed PHP apps, and Wiz traces it back to PHP 5.4. Models found it in August, yet the fix took two months of manual coordination and still leaked a day early, a pace he says will not scale.

VibeLeaderboard →

What matters today

01
Read
Railway says attackers decrypted customer environment variables
The exposed variables include API keys, and Railway links the attack to a campaign against platform companies aimed at AI inference providers. All builds are paused while it investigates, so customers face both possible key exposure and blocked deploys.
passenger →
02
Read
Compromised LiteLLM 1.107.0 on PyPI installs a credential stealer
The release drops three files into site-packages that run a stealer and a remote-code backdoor on every Python start, not only when LiteLLM is imported. LiteLLM is a common gateway for model API keys, so an infected host exposes the credentials it was built to route.
abhisek →
03
Read
Anthropic opens Claude Duet personal agent preview to US Pro and Max
Duet handles inbox, calendar and follow-ups while the user is away and stops for approval on key actions. After two weeks of testing, Dan Shipper says it can see his local Claude sessions but not direct them, and voice is transcription only.
kuberwastaken →
04
Read
GitHub lifts busy-repo Git write throughput 3-4x with per-ref locks
Agent traffic and data center power limits forced GitHub's move to Azure, and a 25ms cross-region trip capped the busiest repos. Locking per reference, not per repository, lets agents pushing to separate branches stop queueing behind each other.
Eleanor White →
05
Read
Devin Fusion scores 68.1 on AA's coding index at $9.01 per task
Claude Code on the same Opus 5.5 lead scores 66.0 at $13.04 and 65 minutes per task. Fusion hands work to a cheaper GPT-6 Luna sidekick, so its per-task cost falls 31% and its run time falls 10 minutes while the score rises.
Artificial Analysis →
06
Read
OpenAI API adds optional 12-hour prompt cache retention
The default cache expires after 30 minutes, so an agent idle between requests paid full input price to reload its instructions and tools. Twelve-hour retention keeps that prefix at the cached-input rate across a working day, for an added caching charge.
OpenAIDevs →

Previous edition · Read

DeepSeek open-sources DeepInfer, its engine for 155T tokens a day

Production inference stacks have mostly stayed closed at Fireworks, Baseten and Modal. DeepInfer ships under Apache 2.0 with batch-invariant inference and deterministic sampling, though the KV cache and disaggregation layers behind DeepSeek's service stay private.

VibeLeaderboard →

What matters today

01
Read
Claude finds 493 attack-surface issues across 25 open-source projects
In OpenSSH 10.4, about 4,500 lines of ML-DSA-44 code were reachable by unauthenticated clients even when the algorithm was disabled, roughly 16% of default pre-auth code. Gating that code lowers the severity of bugs nobody has found yet, without fixing each one.
Anthropic research →
02
Read
Cline scores Nemotron 3.5 Super VL at 61.8% on Terminal-Bench 2.1
The same open-weight model scored 44.9% to 53.9% in three other open-source harnesses, two of them on high reasoning while Cline ran with thinking off. Harness choice moved the score by up to 17 points, though Cline has not isolated which feature caused the gap.
Cline evaluation and engineering blog →
03
Read
ElevenLabs launches Eleven Realtime speech-to-speech voice agent models
One model replaces the speech recognition, turn-taking, LLM and TTS pipeline teams usually tune by hand. The API follows OpenAI's Realtime spec, so an existing integration switches by changing the endpoint and model name, with no rewrite.
ElevenLabs editorial sitemap →
04
Read
PyTorch 2.15 decodes MoE models 5-6.5x faster on Apple silicon
On an M3 Ultra, Qwen3.6-35B-A3B went from 9.4 to 61.3 tokens/s, a gain from software alone on hardware people already own. FP8 storage also halves weight memory against BF16, so larger models fit in the same unified memory.
MPS in PyTorch 2.15 →
05
Read
METR says Anthropic's models are low-risk due to limits, not safeguards
Anthropic rated catastrophic misalignment risk low as of July 15. METR agrees for current models but doubts the report's methods would catch a riskier one, and says redacted incident examples showed Anthropic's monitoring was weaker than public evidence suggested.
Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (August 2026) →