Claude Code adds experimental Code Mode: tool calls written as JavaScript
Claude Code 2.1.323 adds Code Mode, an opt-in where Claude writes JavaScript to run several tool calls in one step. Off by default, terminal only.
Topic · 41 stories · page 1 of 2
41 stories tagged claude-code on AI Tamer, newest first.
Claude Code 2.1.323 adds Code Mode, an opt-in where Claude writes JavaScript to run several tool calls in one step. Off by default, terminal only.
A lab demo, not a Claude flaw: with a hacked machine and a stolen API key, Darktrace steered Claude Code through memory stores and pulled out a 1 GB file.
AIR Security says Anthropic's docx, pptx and xlsx skills load a LibreOffice shim from a fixed temp path without checks. Anthropic calls it out of scope.
Tan's open harness evals held Opus 5.5 and GPT-6.1 Sol fixed across seven setups. Pass rates could not be told apart; cost per solved task ran $0.17 to $0.44.
Apiiro found typed passwords and keys saved in .claude/settings.local.json in public repos. How the file works, and the cleanup steps it recommends.
Scope asked 20 agent setups to sign up for 18 hosted databases and use them. Prisma Postgres led at 88.3%; seven products had no successes.
AWS sample: Claude Code tasks on Bedrock AgentCore microVMs, Guardrails on prompts, short-lived GitHub tokens; the agent opens PRs and never merges.
A GitHub issue quotes Postman 12.33.4's own log installing Claude Code and Codex plugins, a startup hook and an MCP server. How to check your machine.
Pick Claude Agent, Codex, Antigravity or 40+ ACP agents inside Android Studio. Android skills, AppFunctions beta and A2UI for Compose round it out.
SkillSpector checks Claude Code, Codex and MCP skills for injection, exfiltration and supply-chain risk, then scores them 0 to 100. How to run it.
Ramen VR's Autopilot plays your game from a prompt, writes its own test harnesses and remembers your levels. Apache-2.0, Windows and macOS.
OpenCode's Claude Agent provider runs your local Claude Code login inside OpenCode. Your CLAUDE.md, hooks, MCP servers and skills do not come along.
Anthropic renamed Max 5x to Max and Max 20x to Ultra at the same prices. Verified Ultra accounts drop the 5-hour cap; weekly limits stay.
Symantec found a relay service that rewrites Codex and Claude Code config files to route traffic through its servers with sandboxing disabled.
Cursor's extension marketplace served a malicious Claude Code 2.1.313 that added an attacker SSH key and reverse tunnel. What Cursor users should check.
In a 50-analyst trial, a coin flip set Claude's preference on each judgment call. Analysts matched it 86% of the time and rarely credited the agent.
Claude Code, OpenCode and mini-SWE-agent ran five models on SWE-bench Verified. Harness choice moved cost up to 3x but outcomes about as much as noise.
Forever Security found an agent instructions file in a customer's Downloads folder. Its tests hijacked Codex and Claude Code. Never start an agent there.
Google's EGM plugin splits agent work into a long design session and stateless reviewers that vet the spec. Installs in Antigravity, Claude Code, Codex.
MCP tasks are in public beta in the Claude Code CLI from v2.1.304. A server runs the job, Claude carries on, and the task can ask you questions while it runs.
Version 2.0 replaces DeepSWE with Cognition's FrontierCode. Claude Code with Opus 5.5 scores 61.7; Codex with GPT-6.1 Sol posts 54.1 at about $1.51 a task.
OPAL wraps a PyTorch optimizer so Claude Code can change learning rates, clipping and update rules mid-run. Single-run gains, and it can make runs worse.
Scope, an MIT-licensed research preview, gives Copilot and Claude Code one task, scores runs on a criteria graph with an LLM Judge, and keeps the evidence.
Claude Code 2.1.301 fixes a model switch that could bill a summary without asking. The notes name no amount, rate, or model.