AI Coding Assistants

Today's Stories

1. 1

   ActionRail: Open-source framework stops AI agents from executing wrong values

   ToolJet released ActionRail, an open-source framework in public beta that intercepts AI agent tool calls before execution and verifies argument values against live system data—returning allow, hold-for-human, or block decisions. It supports Python applications directly and integrates automatically with LangGraph; v1 is planned for August 2026. Standard permission checks (allowlists, schema validation, policy engines) catch malformed requests but miss well-formed arguments with wrong values—like refunding an already-refunded order or transferring to the wrong customer account. Red-team testing found every model tested executed at least one poisoned value when unprotected (success rates 1.7%–63.3%), but ActionRail blocked 0 of 480 manipulated actions and wrongly blocked 0 of 480 legitimate requests.

   ActionRail grounding rules are testable as executable contracts in CI (Action Tests); sources (SQLite, PostgreSQL, MySQL, HTTP/REST, MCP gateway) stay in your environment and never leave the VPC. The optional control plane and React Console let teams author rules per argument and watch a live activity feed. LangGraph has automatic tool discovery and wrapping; additional framework adapters are planned.

2. 2

   Codeberg blocks AI-generated code uploads

   Codeberg, a code hosting platform, has implemented an automated check to prevent uploads of AI-generated code, targeting bot activity on the service. The move signals concern about machine-generated code flooding open-source repositories, a challenge facing platforms that host developer projects as AI code generation tools become more widespread.

   Users with JavaScript enabled will be redirected automatically; those without may need to manually navigate, though Codeberg indicates the redirect should occur regardless.

3. 3

   Code mode helps smaller AI models call tools reliably

   Instead of training smaller language models to generate properly formatted tool-call signatures, developers can ask them to write Python or TypeScript code that calls APIs or bash commands, which these models already do well from their training data. Anthropic uses this approach to reduce token spending. Smaller language models often fail at generating correctly formatted tool calls, a problem code mode sidesteps by leveraging their existing code-writing ability. This removes the need for expensive retraining while reducing the number of tokens (the data units an AI processes) the system consumes.

   Code execution happens in ephemeral Linux sandbox environments that spin up in under 200ms, adding minimal delay. The tradeoff is the small risk of executing malformed code, which providers mitigate by running each execution in an isolated microVM.

4. 4

   getdebug 0.4.0 adds Python AI-app security patterns, outperforms generic SAST tools

   getdebug CLI 0.4.0 released Python regex-based prefilters for five AI-app anti-patterns (prompt-injection, unsafe-role-merge, pii-in-prompt, unbounded-stream, unsafe-tool-output) alongside existing JS/TS detection. The tool runs deterministically with no LLM call, in milliseconds, at zero cost. In a benchmark against Bandit, Semgrep, and vulnhuntr on real code (Simon Willison's 49-file llm CLI), getdebug surfaced 6 AI-app-specific findings while Bandit returned 1,228 noise alerts (mostly assert_used warnings on pytest) and Semgrep found 3 generic hits with no AI-app coverage. Vulnhuntr selected zero files because it targets web-app entry points, not CLIs. For teams shipping Python apps that call LLMs, getdebug fills a gap the general-purpose security tools don't address.

   getdebug also offers optional local-LLM SAST via Ollama (free, on-device) and hosted Claude-based scanning (paid). A newer version (0.5.5) introduced a stricter detector that now flags one previously-safe fixture as a critical issue, signaling the tool is still tightening its allowlist logic. Benchmark corpus and methodology are open on GitHub (CodeSecBench) and at getdebug.dev/bench; PRs welcome.

5. 5

   AI Toolbox adds Claude Opus 5 support

   AI Toolbox, a tool for working with AI models, now supports Claude Opus 5, Anthropic's latest large language model. Support for Claude Opus 5 expands the range of AI models available to users of AI Toolbox, giving them access to one of the newest AI models for their workflows.

   The tool is available at ai-toolbox.co for users interested in testing Claude Opus 5 integration.