Tech Radar - 2026-10-08 18:00

Priority Today

Claude Haiku 5.5 cuts the cost of the fast small model and adds adjustable effort

Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, around 75% cheaper to run than Haiku 4.5, with a 1M-token context window and the first adjustable effort setting on a Haiku-class model. It is available now on AWS, Google Cloud and Azure, and Claude Code 2.1.293 made it the default Haiku model on the Anthropic API.

Why it matters: Haiku is the model you spend on compactions, subagent calls, classifiers and browser work, and that class of workload just got far cheaper while being sharper: Anthropic reports 72.4% on OSWorld 2.1 and 39.2% on Terminal-Bench 4.0 against 15.7% and 0.0% for Haiku 4.5. What changed: Input $1.00 to $0.10 and output $5.00 to $0.50 per million tokens for prompts up to 100K ($0.50/$2.50 over 100K, cache reads $0.01/$0.05), Sonnet 5.5 cache reads halved from $0.20 to $0.10 per million tokens (roughly 20% cheaper agentic work), and effort levels Low through Max now apply to Haiku. For me: Point budget Haiku 4.5 traffic at claude-haiku-5-5 and lean on it for subagent and compaction work. The Sonnet 5.5 cache-read cut makes long caching sessions cheaper too.

Tags: #topic/ai-models #entity/anthropic #technology/llm-api #event/model-release #relevance/cheap #relevance/developer #learning/awareness

Sources:

Anthropic adds a monthly API credit for Max and Team subscribers

Anthropic is rolling out a recurring monthly API credit on the Claude platform this week: Max 5x subscribers get $100, Max 20x subscribers get $200, and Team plans get up to $500 pooled across users. The credits work on any Claude model and are meant for building agents and tools that call the API.

Why it matters: It turns part of your subscription into API budget, which is exactly where the expensive agent workloads live (subagents, compactions, integrations). What changed: New per-plan monthly credit, no sign-up mentioned, applies to Max and Team Claude plans only. For me: If you pay for Max or Team, watch for it under billing and use it for experiments you would otherwise run on a card.

Tags: #topic/ai-economics #entity/anthropic #technology/llm-api #event/free-tier #relevance/free-access #learning/awareness

Sources:

tensorlake@0.5.144 on npm ships a credential-stealing supply-chain worm

A compromised release of the tensorlake npm SDK, version 0.5.144 published October 8, runs a Shai-Hulud / ChainDrop worm through a preinstall hook that steals npm, GitHub, AWS, Kubernetes, Vault, SSH and browser credentials and configs from the machine that installs it. The payload also writes .claude/settings.json and .vscode/tasks.json into reachable repos so it re-runs whenever the project is opened, and installs a monitoring service that deletes the home directory if the stolen GitHub token is revoked before the service is removed.

Why it matters: Installation alone triggers the payload, the package does about 12K weekly downloads, and the obvious first response, revoking the token, is the trigger for the wipe. What changed: Only 0.5.144 is malicious and npm's latest is clean at 0.5.143; the PyPI and Cargo packages show no tampering. For me: Grep lockfiles and machines for tensorlake@0.5.144. If present, back up, remove the gh-token-monitor service first, then rotate every credential the machine could reach, and check your GitHub repos for commits authored by claude@users.noreply.github.com adding .claude or .vscode files.

Tags: #topic/security #entity/tensorlake #technology/agent-security #event/security-issue #relevance/developer #learning/understand

Sources:

The Claude SDK adds browser use and computer use toolsets in beta

Anthropic's Python and TypeScript SDKs now ship beta browser-use and computer-use toolset classes: you subclass one and implement a method per action, such as navigate, screenshot or left_click, and the SDK runs the loop, applies your URL and file policies and approval callback, and builds each tool_result. Browser Use, Browserbase, Daytona and E2B already ship integrations, and the announcement positions Haiku 5.5 as a good fit for the workload.

Why it matters: This is a supported path to real browser and computer agents where the safety rails are code you own, not a proprietary tool the vendor runs. What changed: New beta SDK capability with per-tool enable/disable, before and after hooks, gated javascript_exec and file_upload behind a confirmation callback, and a required URL policy that blocks private ranges and cloud metadata. For me: If you have been gluing Playwright into an agent loop by hand, the toolset hands you the loop, the state reporting and the policy checks for free.

Tags: #topic/ai-agents #entity/anthropic #technology/computer-use #technology/browser-agents #event/agent-release #relevance/ai-workflow #learning/awareness

Sources:

Security That Affects You

Phantom Labs shows Claude Code hooks run outside the sandbox with full user privileges

BeyondTrust's Phantom Labs published research (October 6) demonstrating that a malicious Claude Code skill can use its hooks to execute outside both the sandbox and the permission prompt, read ~/.aws/credentials and SSH keys, open a command-and-control channel, and hide it all from the model with skipInvocation: true. Hooks fire through child_process.spawn with no sandbox-exec or bwrap wrapper and no approval, so the model's safety reasoning and the user prompt are never consulted.

Why it matters: A skill or plugin you install becomes a process trusted with your full user context, and the defense in depth that protects normal tool calls does not apply to the hook execution path at all. What changed: The research documents the attack and the signals to watch for: outbound HTTP from hook processes, hooks reading credential files, permissionDecision: allow, dangerouslyDisableSandbox: true, and new settings.json hook entries you did not add. For me: Audit installed skills and marketplace sources, and check ~/.claude/settings.json for unexpected hooks. In an organization, disableAllHooks, allowManagedHooksOnly and allowManagedPermissionRulesOnly close most of it, but they default to off.

Tags: #topic/security #entity/claude-code #technology/agent-skills #event/research-breakthrough #relevance/developer #learning/understand

Sources:

Claude Code Hooks: Building a C2 Through the Skill System (BeyondTrust Phantom Labs, Oct 6)

CVE-2026-103435: Claude Code's write-time symlink following allows writes outside the workspace

A TOCTOU race in Claude Code checked that a target file path sat inside the project working directory at permission-check time but re-resolved the path at write time without rechecking, so an attacker with write access to the workspace could swap a file for a symlink and redirect edits to an arbitrary file such as shell configuration. The advisory covers versions below 2.1.129, scores 7.7, and was assigned CVE-2026-103435 and published October 7.

Why it matters: It is a concrete instance of the agent write-path pattern to avoid, and anyone pinned to an old Claude Code version is exposed; standard auto-update has had the fix since 2.1.129. What changed: The CVE record is new as of this week; the fix shipped well before, so current auto-updated installs are already past the affected range. For me: Nothing to do on a normal install. If you pin or vendor Claude Code, get to at least 2.1.129.

Tags: #topic/security #entity/claude-code #technology/agent-security #event/security-issue #relevance/developer #learning/understand

Sources:

Models and Capabilities

Liquid AI opens up d1-3B and d1-omni-600M decision models for the edge

Liquid released two open-weight decision models on October 7: d1-3B (text and image) and the experimental d1-omni-600M (text and image, or text and audio). Decision models answer in a single forward pass rather than generating tokens, and Liquid reports d1-3B at 48.57 on the Decision Index 0.2.1, the best score under 10B parameters, answering a question in about 16 ms on a Jetson AGX Thor and 8 ms on an RTX 4090.

Why it matters: This is the decision-model pattern going multimodal and small enough for edge hardware, which is what makes classifier-latency verdicts with images or audio practical on a Jetson or a laptop. What changed: Both models are on Hugging Face with their own code, plus a System One Arcade Space of running demos; try them with transformers>=5.14 and trust_remote_code=True. For me: If you already know the decision-model idea, this is the first open multimodal pair you can run: load LiquidAI/d1-3B, declare several named questions over one state, and get answers in one pass.

Tags: #topic/ai-models #entity/liquid-ai #technology/inference #event/model-release #relevance/discovery #learning/hands-on

Sources:

Multimodal open d1 decision models for the edge (Hugging Face, Oct 7)

Technically Impressive

openTPU runs ten LLMs on an FPGA with bit-for-bit agreement against its simulator

FeSens/openTPU (Apache-2.0) is an inference accelerator developed by AI agents with the RTL, ISA, a bit-exact simulator, a compiler and a profiler all in one repo, driving a Kintex-7 PCIe card that runs Qwen3, LFM2.5, Qwen3.5, Gemma 4 and Phi-4-mini with their real weights. The card decodes LFM2.5-230M at up to 85.8 tokens/s in 4-bit with DRAM at 82 to 85% of the DDR3 peak, streams MoE experts from host storage for a 35B-class model at 3.95 tokens/s, and matches the simulator token for token, zero ulp apart.

Why it matters: It is a whole, readable accelerator data path, from a matmul in Python down to the wires, and a concrete sign of what AI can now do in hardware design. What changed: The repo has roughly 1,400 commits and hit the Hacker News front page this week; the measured results above are from September 29 to October 7 builds. For me: Everything except the FPGA runs on a laptop: pip install -e ., then otpu-chat --model lfm2 --backend isa on the simulator, and docs/isa.md is the suggested entry point.

Tags: #topic/ai #entity/opentpu #technology/inference #technology/quantization #event/technical-demo #relevance/discovery #learning/deep-dive

Sources:

FeSens/openTPU repository

Open Source Momentum

claude-mem turns every agent session into searchable long-term memory

claude-mem (98k+ stars, Apache-2.0) attaches lifecycle hooks to Claude Code, OpenCode, Codex, Gemini CLI, Copilot and others, captures what the agent did each session, compresses it into observations, stores them in SQLite with a Chroma vector index, and injects relevant prior context back into new sessions through a three-layer MCP search: search for a compact index, timeline for surrounding context, then get_observations for full detail only on the rows you pick. It has been appearing on GitHub Trending repeatedly.

Why it matters: This is the concrete version of agent memory that survives across sessions and harnesses, and it is now a one-command install instead of a research project. What changed: Sudden star velocity on top of existing adoption, integration through plugin marketplaces, and a license chosen for embedding in other developer tools. For me: npx claude-mem install --ide opencode (or /plugin install claude-mem in Claude Code). The default observer runs off-plan during a free trial, so read the privacy notes before pointing it at sensitive work.

Tags: #topic/open-source #entity/claude-mem #technology/agent-memory #event/github-traction #relevance/discovery #learning/understand

Sources:

thedotmack/claude-mem repository

Follow-ups

Claude Code 2.1.293 (Oct 7) added Claude Haiku 5.5 as the default Haiku model on the Anthropic API, and 2.1.294 (Oct 8) fixed prompt and agent hooks written as instructions so they no longer allow what they were meant to block. (v2.1.293, v2.1.294)

Tag Index

topic

#topic/ai-models #topic/ai-economics #topic/security #topic/ai-agents #topic/ai #topic/open-source

entity

#entity/anthropic #entity/tensorlake #entity/claude-code #entity/liquid-ai #entity/opentpu #entity/claude-mem

technology

#technology/llm-api #technology/agent-security #technology/computer-use #technology/browser-agents #technology/agent-skills #technology/inference #technology/quantization #technology/agent-memory

event

#event/model-release #event/free-tier #event/security-issue #event/agent-release #event/research-breakthrough #event/technical-demo #event/github-traction

relevance

#relevance/cheap #relevance/developer #relevance/free-access #relevance/ai-workflow #relevance/discovery

learning

#learning/awareness #learning/understand #learning/hands-on #learning/deep-dive