Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Agentic AI

Clawdbot's week of exposures is a warning for every agent deployment

In the last week of January, Clawdbot, an open-source AI assistant that users host themselves and connect to email and other accounts, went from viral project to security case study. On January 26, the security firm SlowMist warned that hundreds of Clawdbot gateways were exposed to the internet, with API keys, OAuth tokens and private chat histories reachable without credentials, VentureBeat reported. On January 27 the project was renamed Moltbot after Anthropic sent a trademark request over the name’s similarity to “Claude.”

The project had reached 60,000 GitHub stars within weeks. By the end of the week, researchers had documented exposed instances, working prompt-injection attacks, a proof-of-concept supply-chain attack and infostealer malware adding it to their target lists.

/ 01What researchers found

According to VentureBeat, Jamieson O’Reilly, founder of the red-teaming firm Dvuln, found hundreds of exposed instances using Shodan; eight had no authentication and allowed full command execution. Matvey Kukuy, CEO of Archestra AI, used prompt injection in an email to extract an SSH private key from a Clawdbot machine in about five minutes. O’Reilly also uploaded a harmless proof-of-concept “skill” to the community skill registry, inflated its download count, and reached 16 developers in seven countries within eight hours.

VentureBeat reported that Clawdbot stores its memory files, which can include VPN configurations, corporate credentials, API tokens and months of conversation history, in plaintext Markdown and JSON in the user’s home directory, and that the RedLine, Lumma and Vidar infostealers had added it to their target lists.

The security firm Intruder described four active threat areas: exposed credentials, prompt injection (including instances connected to X that leaked private information in response to crafted replies), backdoored community skills, and agents taking actions beyond the user’s intent. It attributed the problem to architecture, saying the platform favours easy deployment over secure defaults and does not sandbox third-party plugins. Its advice to anyone who ran an instance on default settings was to assume compromise.

/ 02Why it matters

It would be easy to dismiss this as a hobbyist tool misconfigured by hobbyists. That misses the point. Clawdbot is a vivid, fast-moving version of what many enterprises are building more slowly: an agent with broad access to a person’s or an organisation’s systems, taking instructions in natural language, and reading untrusted content as part of its work.

Itamar Golan, co-founder of Prompt Security, summed it up for VentureBeat: “It’s an identity and execution problem.” An agent that can read an email and also run commands has merged two things security practice has always kept apart: data and instructions. Any text it reads can become a command.

/ 03The traceability angle

Three design principles follow, and none of them is specific to one product.

  • Separate what the agent knows from what it reads. An agent that treats every incoming document as potential instruction can be steered by anyone who can send it text. An agent that reasons over a curated, governed knowledge base, and treats new input as data to be checked against it, has a smaller attack surface.
  • Constrain actions outside the model. Permissions and policies should be enforced by deterministic checks that the model cannot talk its way past, with the least access needed for each task.
  • Record why each action happened. Intruder’s advice to review activity logs only helps if the logs say what the agent saw, what it concluded and which rule allowed the action. A traceable reasoning chain turns incident response from guesswork into review.

None of this requires abandoning agents. It requires treating them like any other privileged software: scoped credentials, encrypted secrets, reviewed extensions, and logs that make sense to an investigator. The difference with agents is that the decision logic itself is part of what must be logged, because the same tool call can be legitimate or malicious depending on why it was made.

An agent that cannot tell its knowledge from its inbox will eventually take orders from a stranger.

/ 04What to do

If your organisation has staff experimenting with self-hosted agents, find them; VentureBeat’s headline was that infostealers knew about Clawdbot before most security teams did. For sanctioned agent projects, review which systems each agent can reach, where its memory and credentials are stored, and whether its actions can be reconstructed after the fact.

How it unfolded
JAN 26, 2026SlowMist warns of exposed gateways[1]JAN 27, 2026Clawdbot renamed Moltbot[1]JAN 29, 2026This postJAN 26, 2026SlowMist warns of exposed gateways [1]JAN 27, 2026Clawdbot renamed Moltbot [1]JAN 29, 2026This post
Aktiver’s take — Enterprise data

Clawdbot did not fail because its model was weak. It failed because an email, a crafted reply and a downloaded skill all carried the same authority as the owner, and nothing recorded which of them caused a given command.

A graph gives an agent a place to keep that distinction. Every input arrives as a node with a source and a trust level; every action is a node linked to the input that prompted it, the credential it used and the permission that allowed it. A shell command that traces back to an unauthenticated email has no permission edge, so a deterministic check refuses it. After an incident, the investigator’s question becomes a query: which actions in the last week trace to external content, and which secrets did they touch. Agents built in Knowledge Builder reason over the governed graph rather than raw inbox text, and running them behind the firewall keeps their memory off the public internet.

How we’d model it

Agent actionInbound messageUntrustedAPI tokenPermissiontriggeredByhasTrustLevelusesCredentialauthorizedByAgent actionInboundmessageUntrustedAPI tokenPermissiontriggeredByhasTrustLevelusesCredentialauthorizedBy
  • Agent action —triggeredBy→ Inbound message
  • Inbound message —hasTrustLevel→ Untrusted
  • Agent action —usesCredential→ API token
  • Agent action —authorizedBy→ Permission

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading