Tech Radar - 2026-10-14 18:00

Priority Today

Check Point found two high-severity LiteLLM flaws, fixed in 1.83.9

Check Point Research published an assessment of LiteLLM 1.77.7 on October 13, 2026 that found two vulnerabilities, CVE-2026-35029 and CVE-2026-59823, both resolved in LiteLLM 1.83.9. The access-control flaw left the admin-only /config/update endpoint open to any authenticated user, which enabled arbitrary file read (via UI_LOGO_PATH and /get_image), admin takeover by overwriting UI_USERNAME/UI_PASSWORD, and remote code execution by registering pass-through handlers that load Python from S3 or GCS; the second flaw is an SSRF in the Clientside LLM Credentials feature that can redirect calls and leak provider API keys.

Why it matters: A gateway sits in front of every provider key, so one low-privilege account could reach the keys and the runtime configuration that all traffic depends on.

What changed: Two disclosed CVEs with fixes shipped in 1.83.9 (NVD rates CVE-2026-35029 CVSS 3.1 8.8; CVE-2026-59823 as moderate). Check Point notes LiteLLM is often embedded in other platforms, so it may be running where it was never deployed directly.

For me: Upgrade to 1.83.9 or later, rotate every provider key the proxy held, and restrict /config/update to an admin network. Check for LiteLLM inside other tools before assuming you do not run it.

Tags: #topic/security #topic/ai-agents #entity/litellm #technology/agent-security #technology/llm-api #event/security-issue #relevance/developer #relevance/self-hosting

Sources:

NVIDIA open sourced the full Run:ai GPU orchestration platform for Kubernetes

On October 13, 2026 NVIDIA released the complete Run:ai platform as open source under its DSX AI Factory GitHub organization, including four components that can also be adopted independently: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. KAI Scheduler is Apache-2.0 and already onboarded as a CNCF Sandbox project, and it can run alongside an existing cluster without the rest of the platform.

Why it matters: GPU scheduling and fair-share on shared clusters has mostly been vendor-locked; the scheduler in particular is now upstreamable and inspectable, which matters for teams running training and inference on the same hardware.

What changed: Production Run:ai code, not just the scheduler, is now public and consumable component-by-component.

For me: If you run Kubernetes with GPUs, evaluate KAI Scheduler on its own before adopting the whole platform.

Tags: #topic/open-source #topic/self-hosting #entity/nvidia #entity/kubernetes #technology/containerization #event/open-source-release #relevance/self-hosting #relevance/infrastructure

Sources:

Liquid AI released LFM2.5-24B-A2B, a 24B open-weight MoE with 2B active parameters

Liquid AI released LFM2.5-24B-A2B on October 13, 2026, a 24B-parameter mixture-of-experts reasoning model with about 2B active parameters per token, a 131K context window and a 128K vocabulary, post-trained for tool calling and multi-step agentic work on-device. The base and post-trained checkpoints are open-weight on Hugging Face with native llama.cpp, MLX and vLLM support.

Why it matters: A capable agentic model small enough to run locally makes private, offline tool-calling agents practical on a laptop or a single GPU.

What changed: A new largest-and-final model in the LFM2.5 family, open-weight, with a doubled 128K vocabulary and pretraining scaled from roughly 17T to 33T tokens. Liquid AI reports it is competitive with Qwen3.6-35B-A3B on five of eight agentic benchmarks and runs up to 1.7x faster than comparable MoEs on-device (about 184 tokens/s on an Apple M5 Max).

For me: Try the MLX or GGUF builds for a local agent; treat the benchmark comparisons as vendor-reported until you reproduce them.

Tags: #topic/ai-models #topic/ai-agents #entity/liquid-ai #technology/inference #event/model-release #relevance/developer #relevance/ai-workflow #learning/understand

Sources:

Roo Code's command approval can be bypassed with a bare redirect

CVE-2026-110202 (VulnCheck, published October 13, 2026) covers Roo Code through 3.54.0, where getCommandDecision() can auto-approve a denied shell command prefixed with a bare >& redirect. Because the allowlist matches the redirect target (for example echo) while the shell still runs the denied command, indirect prompt injection is enough to execute arbitrary commands on the developer machine.

Why it matters: It is the same class of failure as several recent agent gate bypasses: the approval check parses the command the wrong way and the shell expands it differently, so the guardrail never sees what actually runs.

What changed: A published CVE against a widely used VS Code coding agent; no fixed release version is stated in the advisory, so mitigation is behavioral until a patch lands.

For me: If you use Roo Code, stop auto-approving shell commands and do not run it against untrusted repositories or prompts until a fix is available.

Tags: #topic/security #topic/ai-coding #entity/roo-code #technology/agent-security #event/security-issue #relevance/developer #relevance/ai-workflow #learning/understand

Sources:

Cloudflare's Agents SDK added a persistent browser that keeps tabs and logins between turns

Cloudflare shipped a persistent browser tool in its Agents SDK (changelog October 13, 2026): a remote Browser Run session stays open between tool calls, so tabs, cookies and login state survive into a later turn while the browser is alive. The agent drives it over the Chrome DevTools Protocol, the Durable Object reconnects after hibernation, and the default idle timeout is 10 minutes (Agents SDK 0.28.0 or later).

Why it matters: Multi-turn browser tasks no longer require building your own session store or re-authenticating every turn, which is the usual blocker for agents that log into a site and return to it later.

What changed: A new browserTool with durable, resumable sessions, plus a Live View handoff so a person can complete a login, MFA or CAPTCHA in the same browser.

For me: Good fit for agents that check a dashboard or fill a form and come back; design around the 10-minute idle window and treat a restarted browser as lost state, not durable storage.

Tags: #topic/ai-agents #topic/developer-tools #entity/cloudflare #technology/browser-agents #event/agent-release #relevance/developer #relevance/ai-workflow

Sources:

AI Coding and Agents

Claude Code 2.1.299 compacts a long idle conversation before its prompt cache expires

Claude Code 2.1.299 (October 14, 2026) adds idle compaction: a conversation over 200K tokens left idle for about 54 minutes is compacted before the one-hour prompt cache expires, so the next message does not reprocess the whole context. It also adds CLAUDE_CODE_DISABLE_FILE_MENTIONS, which gives the model the paths of @ file mentions instead of the files' contents, and CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS to change the 200K minimum.

Why it matters: It targets the cost and latency spike that hits when a long session's cache lapses between messages and the whole conversation is resent.

What changed: New context-management behavior plus two environment variables; "idleCompaction": false turns it off.

For me: Nothing to configure on recent versions; if you deliberately keep a large working set in context, be aware it can now be compacted underneath an idle session.

Tags: #topic/ai-coding #entity/claude-code #technology/context-engineering #event/tool-release #relevance/developer #learning/awareness

Sources:

Infrastructure and DevOps

AIBrix v0.8.0 adds a ModelClaim controller that packs several models onto one GPU

AIBrix v0.8.0 (released October 9, blog October 12, 2026) is a Kubernetes-native inference stack whose new ModelClaim controller shares a single GPU across multiple models: it uses kvcached for elastic KV-cache allocation, vLLM sleep mode to swap idle engines to host memory, and a per-Pod memory ledger so no model's reservation is double-counted. The release is 443 commits from 104 contributors since v0.7.0.

Why it matters: It targets the common case of many long-tail, low-traffic models that would otherwise each hold a whole GPU idle.

What changed: An experimental model-lifecycle control plane alongside inference-aware routing. ModelClaim supports vLLM only and is marked experimental.

For me: Worth testing on a low-traffic model pool if you run vLLM on Kubernetes; expect to tune the per-GPU footprint and KV-floor figures by hand.

Tags: #topic/devops #topic/self-hosting #entity/aibrix #entity/kubernetes #technology/inference #event/tool-release #relevance/infrastructure #relevance/self-hosting

Sources:

Databases

Tiger Data's PRISM index runs vector search in PostgreSQL with partitioning and RaBitQ

Tiger Data launched pg_vectorsearch, an alpha PostgreSQL 18 extension whose PRISM (Partitioned Routing Index for Similarity Matching) index replaces HNSW's graph with a hierarchical centroid tree and RaBitQ-quantized vectors stored in SIMD-optimized posting lists. On Tiger Data's 100M-vector Cohere benchmark it reports 6 ms p50 at 95% recall on a 128 GB machine versus HNSW's 8 ms on 512 GB, and a 1h40m build versus 42h; it indexes existing pgvector vector/halfvec columns in place.

Why it matters: Vector search is the cost center in most RAG stacks, and a smaller index that fits in far less RAM makes a separate vector database harder to justify.

What changed: A new PostgreSQL-native index that is pgvector-compatible but not pgvector-dependent; pgvectorscale moves to maintenance mode. Numbers are first-party and the extension is alpha.

For me: Try PRISM on a copy of a real pgvector workload and measure recall yourself; do not put the alpha in production, and expect to drop and recreate indexes across prereleases.

Tags: #topic/databases #entity/tiger-data #entity/postgresql #technology/vector-search #technology/quantization #event/open-source-release #relevance/developer #learning/understand

Sources:

Open Source Momentum

Ruflo's agent meta-harness wraps Claude Code with coordinated agent swarms

Ruflo (formerly Claude Flow) is an MIT-licensed agent meta-harness for Claude Code and Codex: it adds coordinated swarms of specialized agents, self-learning and HNSW vector memory, zero-trust federation across machines, and an MCP server exposing hundreds of tools, installed with one npx ruflo init. Its GitHub repository shows around 68,000 stars, and the npm package around 60,000 weekly downloads.

Why it matters: It is a large, actively used alternative to hand-rolling agent orchestration, memory and multi-agent coordination on top of a single coding agent.

What changed: A mature, fast-moving project that keeps appearing on trending lists; it modifies .claude/ configuration and installs hooks, so it is not a drop-in.

For me: Evaluate it in a throwaway repo first. The star and download counts signal adoption, not that the coordination model is right for your workflow.

Tags: #topic/open-source #topic/ai-agents #entity/ruflo #technology/agent-orchestration #technology/agent-memory #event/github-traction #relevance/developer #learning/awareness

Sources:

New Concepts

mcp-server-cards

MCP Server Cards (SEP-2127, Final on October 6, 2026) is an Extensions Track spec that lets a remote MCP server publish a static, cacheable card describing its identity, transport endpoints and supported protocol versions before any client connects, served at /server-card by default. It deliberately omits the server's tools, resources and prompts, which stay runtime-only, and domain-level discovery moves up to an AI Catalog at /.well-known/ai-catalog.json.

Why it matters: Pre-connection discovery lets IDEs and registries find and configure MCP servers without a configured transport or a live session, and the spec is explicit that card contents are advisory and must not drive access-control decisions.

What changed: The extension reached Final, giving clients and registries a standard out-of-band discovery format alongside the runtime server/discover RPC.

For me: If you run a remote MCP server, publish a card at /server-card next to your endpoint; when consuming cards, allowlist trusted catalog domains and require approval before autoconfiguring to an endpoint on a different domain.

Tags: #topic/ai-agents #entity/modelcontextprotocol #technology/mcp #technology/agent-orchestration #event/research-breakthrough #relevance/developer #learning/understand

Sources:

Tag Index

topic

#topic/ai-agents #topic/ai-coding #topic/ai-models #topic/databases #topic/developer-tools #topic/devops #topic/open-source #topic/security #topic/self-hosting

entity

#entity/aibrix #entity/claude-code #entity/cloudflare #entity/kubernetes #entity/liquid-ai #entity/litellm #entity/modelcontextprotocol #entity/nvidia #entity/postgresql #entity/roo-code #entity/ruflo #entity/tiger-data

technology

#technology/agent-memory #technology/agent-orchestration #technology/agent-security #technology/browser-agents #technology/containerization #technology/context-engineering #technology/inference #technology/llm-api #technology/mcp #technology/quantization #technology/vector-search

event

#event/agent-release #event/github-traction #event/model-release #event/open-source-release #event/research-breakthrough #event/security-issue #event/tool-release

relevance

#relevance/ai-workflow #relevance/developer #relevance/infrastructure #relevance/self-hosting

learning

#learning/awareness #learning/understand