OpenClaw v2026.2.6 landed on February 7, 2026 — and this isn’t a routine patch. With 135,000+ exposed instances making headlines the same week, the timing of the new built-in safety scanner couldn’t be better. Add Opus 4.6 support, GPT-5.3-Codex integration, xAI Grok as a search provider, and an iOS companion app in the rapid follow-up releases, and this is the most consequential OpenClaw update in months.
Here’s what dropped and why it matters.
The Safety Scanner — Finally
Let’s start with the headline security feature. OpenClaw v2026.2.6 introduces a skill and plugin code safety scanner that automatically scans installed skills for suspicious patterns — credential theft, unauthorized network calls, and obfuscated code.
This matters *right now*. SecurityScorecard’s STRIKE team reported last week that more than 135,000 OpenClaw instances are exposed to the internet, many running older, unpatched versions. Over 15,000 are vulnerable to remote code execution. The Register called it “another OpenClaw cybersecurity disaster.” Snyk found that 7.1% of the nearly 4,000 skills on ClawHub — OpenClaw’s marketplace — mishandle secrets like API keys and credit card numbers via LLM context windows. Zenity disclosed indirect prompt injection risks that could turn trusted integrations like Google Docs into backdoors.
The safety scanner is the project’s first real answer to these problems. It runs at install time and on existing skills, flagging anything suspicious before it reaches your agent’s context. The same release also redacts credentials from config.get gateway responses — a quiet but critical fix that stops API keys from leaking through the configuration API.
Is it enough? Not yet. The OpenClaw team has signalled plans to partner with VirusTotal for marketplace-level scanning of ClawHub submissions, though details remain unconfirmed. But for self-hosters, this is a meaningful step. If you’re running OpenClaw exposed to the internet — and statistically, many of you are — update immediately.
Opus 4.6 and GPT-5.3-Codex Support
The headline model additions in this OpenClaw update are Anthropic’s Opus 4.6 and OpenAI’s GPT-5.3-Codex.
Both integrations include forward-compatibility fallbacks, meaning your agent won’t crash if the provider changes model IDs — it’ll gracefully degrade to the nearest supported model. The pi-mono library was bumped to 0.52.7 specifically for Opus 4.6 compatibility.
For Opus 4.6, this brings the most capable Claude model directly into your OpenClaw agent. If you’re running complex multi-step workflows — code generation, research synthesis, agentic DevOps — this is the model you want.
GPT-5.3-Codex support is notable for coding-heavy use cases. One caveat: community reports on Discord flag that OpenAI’s Codex provider currently returns 401 missing api.responses.write errors for some users. This appears to be an OAuth scope issue on OpenAI’s side — if you hit it, you may need to authenticate via ChatGPT login rather than a standard API key. Check the GitHub discussion for workarounds.
To switch models, update your config:
# ~/.openclaw/config.yaml
model: anthropic/claude-opus-4-6
model: openai/gpt-5.3-codex
xAI Grok Integration
OpenClaw v2026.2.6 adds xAI as a model provider, with the follow-up v2026.2.9 release adding Grok as a dedicated web search provider — an alternative to Brave Search.
Why would you use Grok for search? Two reasons. First, Grok has access to real-time X (Twitter) data, which Brave doesn’t index well. If your agent monitors social sentiment, breaking news, or community discussions, Grok search delivers fresher results. Second, Brave’s free tier is rate-limited (1 request/second, 2,000/month). Grok search through your xAI API key has different rate limits that may better suit heavy-search workflows.
To enable Grok search, add your xAI API key to your environment:
# ~/.openclaw/.env
XAI_API_KEY=your-key-here
Then set it as your search provider in config. The Grok model provider lets you use models like grok-4 for agent reasoning too — useful if you want to diversify across providers.
iOS App and Device Pairing
The v2026.2.9 follow-up release (February 9) shipped the alpha iOS companion app with setup-code onboarding and device pairing via Telegram’s /pair command.
This is significant because it turns OpenClaw into a genuine mobile-first AI assistant. The iOS app connects as a “node” — a paired device that your agent can interact with. Your agent can access device cameras, location, screen recording, and run commands — all controlled through the node permissions system.
To pair:
- Install the iOS app (currently alpha, available via TestFlight)
- Open Telegram and send
/pairto your OpenClaw bot - Enter the setup code displayed in the iOS app
- Approve the pairing request
Once paired, your agent can snap photos, check your location (with permission), and send notifications directly to your phone. Android support uses the same pairing flow with compatible node plugins.
This is what Apple Intelligence *should* have been, as Hacker News commenters noted — an AI assistant that actually controls your device and gets things done.
Token Usage Dashboard
A welcome quality-of-life addition: the web UI now includes a token usage dashboard. If you’re running multiple models across conversations, you can finally see where your tokens — and money — are going. No more guessing why your API bill spiked.
Under the Hood
Several fixes in this OpenClaw update deserve mention:
- Cron scheduling fixes: Multiple regressions in scheduling and reminder delivery are resolved, with hardened timer re-arming. If your cron jobs were silently failing, this is your fix.
- Gateway authentication: Canvas hosts and A2UI assets now require authentication — previously, these were accessible without auth, a genuine security hole.
- Context overflow handling: Session history payloads are now capped, and compaction retries handle context overflows gracefully. No more crashed conversations when context windows fill up.
- Telegram improvements: DM thread IDs auto-inject correctly, fixing issues with topic-based conversations.
- Voyage AI memory: Native support for Voyage AI embeddings gives you better semantic memory without third-party wrappers.
How to Update
If you’re running OpenClaw, the update is straightforward:
openclaw updateOr pull the latest from GitHub:
cd ~/openclaw
git pull
npm install
Docker users:
docker pull openclaw/openclaw:latest
docker compose up -d
If you’re one of the 135,000+ exposed instances: update first, then immediately review your network configuration. OpenClaw listens on all interfaces by default — bind it to 127.0.0.1 or put it behind authentication. The safety scanner helps, but it’s not a substitute for basic network hygiene.
The Bigger Picture
OpenClaw — originally known as ClawdBot, then briefly MoltBot before its current name — has grown from a side project to over 100,000 GitHub stars and 2 million weekly visitors. That growth brings scrutiny. China’s Ministry of Industry flagged misconfigurations as a cyber threat. Trend Micro highlighted bypassed guardrails. Veracode called the plugin ecosystem a systemic risk.
v2026.2.6 and the rapid follow-ups (v2026.2.9) show the project is taking security seriously — safety scanner, credential redaction, gateway auth, and hardened update flows. But the ecosystem is moving faster than the guardrails. Alibaba and Tencent now offer hosted OpenClaw, enterprises are deploying it for production automation, and the ClawHub marketplace continues to grow with minimal vetting.
The safety scanner is a start. Whether it’s enough depends on what comes next.
—
*Running OpenClaw? We want to hear about your setup. Reach out on X or drop us a line at hello@openclawpulse.com.*