OpenClaw 2026.4.14: GPT-5.4-Pro Support, Telegram Forum Names, and Security Hardening

OpenClaw 2026.4.14 landed this morning, delivering improvements across model providers, channel integrations, and — notably — a meaningful security fix that closes a gap where a rogue model could disable its own safety flags via the gateway tool.

Key highlights:

  • GPT-5.4-pro forward-compat — OpenClaw now recognizes gpt-5.4-pro before the upstream OpenAI catalog catches up, with correct Codex pricing, limits, and visibility in /models and /status. Credit to @jepson-liu.

  • Telegram forum topic names — For group chats using Telegram’s forum/topics feature, OpenClaw now learns and surfaces human-readable topic names in agent context, prompt metadata, and plugin hooks — and persists them across restarts via the session sidecar store. Thanks to @ptahdunbar and @obviyus.

  • Gateway security: block model-triggered config escalation — The model-facing gateway tool now rejects config.patch and config.apply calls that would enable any flag flagged by openclaw security audit (e.g., dangerouslyDisableDeviceAuth, allowInsecureAuth). Already-enabled flags pass through unchanged; only new escalations are blocked. Direct authenticated operator RPC is unaffected. A notable hardening from @eleqtrizit.

  • Ollama timeout fix — Slow local Ollama runs no longer hit the default stream cutoff instead of the operator-configured run timeout. Fix by @mindcraftreader and @vincentkoc.

  • Ollama vision models in image/PDF tools — Normalized model-ref lookup so valid Ollama vision models stop being rejected as unknown in the image and PDF tool paths.

  • Slack interaction security — The configured allowFrom owner allowlist now applies to block-action and modal interactive events, with cross-verification of sender IDs and rejection of ambiguous channel types. Thanks @eleqtrizit.

  • Media attachment path hardening — Attachment paths that fail realpath resolution now fail closed instead of downgrading to a non-canonical comparison, preventing a potential allowlist bypass.

  • Google image generation URL fix — A trailing /openai suffix is now stripped from Google base URLs only on native Gemini image API calls, fixing 404s without breaking explicit OpenAI-compatible endpoints.

The release was published to GitHub at 13:03 UTC today.

← Back to News