An AI tool with a funny name has caused quite a commotion as of late—including some allegations of machine consciousness—so here is a breakdown on OpenClaw.
Launched in November 2025, OpenClaw is an open-source, autonomous artificial intelligence (AI) agent that was made to run locally on your own computer, allowing it to manage tasks, interact with applications, and read and write files directly. It acts as a personal digital assistant, integrating with chat apps like WhatsApp and Discord to automate emails, scan calendars, and browse the internet for information.
OpenClaw was formerly known as ClawdBot, but the project brushed up against the large AI developer Anthropic, because of its own tool named “Claude.” In response, OpenClaw’s developer quickly renamed the project to “Moltbot,” which brought impersonation campaigns from cybercriminals. The trademark trouble and the abuse that followed put a dent in OpenClaw’s reputation.
Another dent followed when Hudson Rock published an article about the first observed case of an infostealer grabbing a complete OpenClaw configuration from an infected system, effectively looting the “identity” of a personal AI agent rather than just browser passwords.
The case underlines an impending danger—and not just for OpenClaw, but for other AI agents as well. Infostealers are starting to harvest not just credentials but entire AI personas plus their cryptographic “skeleton keys,” turning one compromised agent into a pivot point for full‑blown account takeover and long‑term profiling.
As I stated before in a broader context, adversaries are starting to target AI systems at the supply‑chain level, quietly poisoning training data and inserting backdoors that only surface under specific conditions. OpenClaw sits squarely in this emerging risk zone: open source, moving fast, and increasingly wired into mailboxes, cloud drives, and business workflows while its security model is still being improvised.
At this stage of its development, treating OpenClaw as a hardened productivity tool is wishful thinking, since it behaves more like an over‑eager intern with an adventurous nature, a long memory, and no real understanding of what should stay private.
Researchers and regulators have already documented prompt injection risks, log poisoning, and exposed instances that hand attackers plaintext credentials or tokens via poisoned emails, websites, or logs that the agent dutifully processes.
How to use OpenClaw safely
For anyone thinking about using OpenClaw in production, the bigger picture is even less comforting. OpenClaw runs locally but is designed to be adventurous: it can browse, run shell commands, read and write files, and chain “skills” together without a human checking every step. Misconfigured permissions, over‑privileged skills, and a culture of “just give it access so it can help” mean the agent often sits at the center of your accounts, tokens, and documents, with very few guardrails.
In fact, an employee at Meta who works in AI safety and alignment recently shared on the social media platform X that she was unable to prevent ClawBot from deleting a major portion of her email inbox.
Further, the Dutch data protection authority (Autoriteit Persoonsgegevens) warned organizations not to deploy experimental agents like OpenClaw on systems that handle sensitive or regulated data at all, flagging the combination of privileged local access, immature security engineering, and a rapidly growing ecosystem of dubious third‑party plugins as a kind of Trojan horse on the endpoint.
Microsoft provided a list of recommendations in this field that make a lot of sense. They are not specifically aimed at OpenClaw, but provide a conservative baseline for self‑hosted, Internet‑connected agents with durable credentials. (If these recommendations feel overly technical, it’s because safely using an AI agent with broad access is still an experimental and technical process.)
- Run OpenClaw (or similar agents) in a sandboxed VM or container on isolated hosts, with default‑deny egress and tightly scoped allow‑lists.
- Give the runtime its own non‑human service identities, least privilege, short token lifetimes, and no direct access to production secrets or sensitive data.
- Treat skill/extension installation as introducing new code into a privileged environment: restrict registries, validate provenance, and monitor for rare or newly seen skills.
- Log and periodically review agent memory/state and behavior for durable instruction changes, especially after ingesting untrusted content or shared feeds.
- Understand and provide for the event where you may need to nuke‑and‑pave: keep non‑sensitive state snapshots handy, document a rebuild and credential‑rotation playbook, and rehearse it.
- Run an up to date real-time anti-malware solution that can detect information stealers and other malware.
We don’t just report on threats—we remove them
Cybersecurity risks should never spread beyond a headline. Keep threats off your devices by downloading Malwarebytes today.
