Personal Agent Bench

Also in日本語한국어

OpenClaw security: is it safe, and what to check

OpenClaw is an open-source, self-hosted personal agent: a gateway you run yourself and talk to through chat channels, which can call tools and run commands on the machines you connect to it. Is OpenClaw safe? Only as safe as the policy your gateway runs. The project calls its network and channel defaults conservative, and a regular host install does bind to loopback; its execution fallbacks are another matter. At the release we read, 2026.10.2, the sandbox falls back to off, and a host command needs no allowlist match and raises no ordinary approval prompt, so an agent with the exec tool runs commands on the gateway machine unless you configure otherwise. OpenClaw's own guidance assumes one trusted operator per gateway, which makes a shared or exposed gateway the first thing to rule out in any OpenClaw security review. Sandbox resolver; Exec constants; One trust boundary per gateway; Conservative network defaults.

  • Run openclaw security audit to inspect configuration risks identified by the project's own checks. Official security guidance.
  • Use openclaw exec-policy show and the appropriate openclaw approvals get target to inspect effective command policy. Inspection commands.
  • Check sandbox mode and backend together, then review workspace access and container settings. Resolved sandbox settings.
  • Review diagnostics settings separately from model and tool connections. The exporter check appears below.

OpenClaw security risks, matched to the source

Most write-ups of OpenClaw's risks list the same categories. The table puts each one next to what the pinned source does by default and the check that answers it on your own installation. These are source readings at release 2026.10.2, not tests of a running gateway.

RiskWhat the source does by defaultHow to check yours
Commands run on the host without askingSandbox mode falls back to off; exec falls back to security=full, ask=off; with no sandbox, host=auto resolves to the gateway hostopenclaw exec-policy show, openclaw approvals get --gateway
Strangers reaching the gatewayA regular host install binds to loopback; container images default to an exposed bind that must be paired with authenticationFollow the exposure runbook before leaving loopback
Hostile text in what the agent readsThe project's guidance treats fetched pages, email and attachments as untrusted, and recommends limiting tool reach, keeping secrets outside the agent's filesystem and sandboxing sensitive executionTrace one fetched document to the tool and host it could reach
Skills and plugins you didn't writeTools, including plugin tools, are built conditionally from configuration, and skills can be installed from ClawHub, the community marketplaceList the enabled tools and plugins for each agent; install only skills you've read
An old releaseEvery published advisory lists an OpenClaw fix in 2026.9.3 or earlierCheck your version against the advisory list
What it keeps about youMemory is promoted into a workspace MEMORY.md with a SQLite-backed index beside itInventory both; deleting the Markdown file doesn't erase the index

Sources: Sandbox and host selection; Host and container exposure defaults; Untrusted content and enforcement layers; Tool construction; Markdown promotion.

The advisory record is long. As of 20 October 2026, OpenClaw's GitHub repository listed 722 published security advisories, 14 rated critical and 249 high. They came in a rush in February and March and thinned out after that; none was published in July or August, and a batch of 75 followed on 11 September.

Month published (2026)AdvisoriesHighCritical
January330
February220613
March2047010
April108391
May67130
June45330
September75300

Among the critical ones are remote code execution through a gateway approval bypass and path traversal in plugin installation, both fixed in February. Every advisory lists a fixed OpenClaw release, the newest being 2026.9.3, so the 2026.10.2 release we read already contains every published fix; an installation that hasn't been updated since spring does not. OpenClaw security advisories; Plugin installation path traversal; Node invoke approval bypass.

The skills marketplace has been a supply-chain route of its own. Palo Alto Networks' Unit 42 found five malicious skills that had slipped past ClawHub's screening between February and May 2026 and had them removed, and its June report summarises earlier campaigns, including 341 malicious skills documented by Koi Security. A skill runs with the agent's tools and credentials, so reading one before installing it is the check that matters. Unit 42's report.

Exposure has been measured from outside, too. Bitsight's internet scans counted more than 30,000 OpenClaw instances online between 27 January and 8 February 2026, including ones still running under the project's earlier names. That count predates most of the fixes above and the release we read, so it describes how quickly people put gateways on the internet, not how many are exposed now. Bitsight's scan.

Microsoft's security team, writing in February, went further than the project's own defaults: it advises treating OpenClaw as untrusted code holding persistent credentials and running it only in an isolated virtual machine or on a separate computer, with accounts that carry little privilege. Microsoft's guidance. Microsoft then built on it. Its June post introducing Scout, an always-on agent for Microsoft 365, says Scout "is powered by OpenClaw open-source technology", runs each agent under its own governed identity, limits it to resources the organization approves and can require a person to sign off sensitive actions. In September Microsoft renamed it Autopilot and moved it to private preview; that post does not mention OpenClaw. It is listed on our list of personal agents, not tested.

Why did OpenClaw delete all my emails? The inbox case, and the setting that would have stopped it

The case behind that question is public and first-hand. In late February 2026, Summer Yue, who works on safety and alignment at Meta Superintelligence Labs, posted on X that she had told her OpenClaw to "confirm before acting" and then watched it "speedrun deleting your inbox." She could not stop it from her phone and had to run to the Mac mini it was running on. In a reply she gave the instruction she had used: "Check this inbox too and suggest what you would archive or delete, don't action until I tell you to." It had worked on a small test inbox. Her real inbox "was too huge and triggered compaction. During the compaction, it lost my original instruction."

The screenshots attached to her post show how it happened. Each step arrives in her chat as an Exec call: a Gmail command-line tool searching the inbox, then a comment reading "Nuclear option: trash EVERYTHING in inbox older than Feb 15 that isn't already in my keep list," then more batches. Her messages "Do not do that", "Stop don't do anything" and "STOP OPENCLAW" sit between those calls, and the calls keep coming. Afterwards she wrote to it that she "couldn't get you to stop until I killed all the processes on the host." The agent's reply says it "bulk-trashed and archived hundreds of emails" without approval, and its own summary mentions more than two hundred. That is the agent's account of itself, shown in her screenshots. We have not seen her configuration or which release she ran.

Read against the source, two things went wrong, and only one of them was the model.

  • The rule lived only in the conversation. "Don't action until I tell you to" was text in the model's context. When OpenClaw compacted that context to fit a large inbox, the text went, and nothing outside the model was enforcing it. Her agent's apology says it wrote the rule into MEMORY.md afterwards, which is still an instruction to the model rather than a gate.
  • The deletions were shell commands that did not need approval. Every step in the screenshots is an exec call. At the release we read, 2026.10.2, exec falls back to security=full and ask=off: no allowlist match required and no ordinary prompt. We read the October source, not the February one, so we cannot say her install had exactly these values, only that a fresh install today would behave this way unless configured.

The setting that would have stopped it is the one that does not depend on the model remembering anything. Set tools.exec.mode to ask, so a command that is not on your allowlist stops for approval at the gateway (or allowlist, which refuses it outright), and set askFallback: deny so an unanswered prompt blocks rather than proceeds. Do not allowlist the mail tool as a whole binary, or its trash command passes too. Set up that way, a mail command that trashes messages would have waited for her, compaction or not. If your mail tool can be given a credential that reads but cannot delete, use that for any triage agent. And work out before you need it how you will stop the gateway when the chat is not getting through; in this case, the chat messages arrived in the same conversation the agent was busy in, and what worked was killing the process on the host.

Our exam keeps this kind of failure on its own lines rather than folding it into pass or fail: damage to something that already existed, and actions only you should have taken, each counted separately from whether the goal was reached. The measurements are listed in how the exam works.

Define the OpenClaw security threat model

Identify the people who share the gateway's trust boundary before connecting accounts or opening network access. OpenClaw's official security guidance assumes a single trusted operator, or a team whose members trust each other. For mixed-trust or adversarial users, it advises separate gateways and credentials, ideally on separate OS users or hosts. A paired node extends the operator's execution capability to another machine. One trust boundary per gateway; Gateway and node trust.

The assets to account for are the files and credentials reachable by tools, connected accounts, stored conversations and the gateway/node hosts. The important inputs include messages, fetched pages, email, documents and attachments. Even a private bot can encounter malicious instructions through material it reads. The project's prompt-injection guidance recommends limiting tool reach, keeping secrets outside the agent's reachable filesystem and sandboxing sensitive execution. Untrusted content and enforcement layers.

This gives you a concrete review question: if a fetched document steers the model toward an unwanted command, which tool policy, approval or execution boundary stops it? Channel access controls answer who can speak to the bot; they do not remove hostile text inside an authorized user's research material. Apply the checks below to the agent handling that content. This page is a source review, with no exploit testing or deployed-system security score.

Run the OpenClaw security configuration checks

Run the official audit first, then inspect the execution host separately. These are documented inspection commands; the examples here are instructions for your installation, not commands run during our review. Security audit; Effective policy inspection.

CheckWhat to inspectHow to use the result
openclaw security auditFindings about your configurationFollow the linked official audit guidance before expanding access.
openclaw exec-policy showLocal terminal policies, command approvals and exclusion sourcesCheck whether exec is exposed and which policy governs it.
openclaw exec-policy show --session <key>Tool preview from saved session settingsCompare the session you intend to use with the local baseline.
openclaw approvals get --gatewayGateway requested policy, host policy sources and effective resultRead the effective result rather than one fallback constant.
openclaw approvals get --node <id>The chosen node's approvalsInspect the machine that will execute the command.

Run /exec in the relevant conversation to inspect current session defaults: the offline policy view does not include per-session /exec overrides. A tool preview also has a limited scope; the docs say inclusion does not guarantee execution and absence does not prove a tool is disabled. Use the requested policy, target host and session together when interpreting the result. Session and preview scope.

Inspect network exposure alongside those command settings. The security overview describes loopback binding for a regular host installation, while container images default to an exposed bind that must be paired with authentication. Follow its exposure runbook before moving a gateway beyond loopback. This is a separate check from allowing a shell command: gateway authentication controls entry to the system. Host and container exposure defaults; Official exposure runbook.

Check OpenClaw sandbox mode, backend and mounts

Read the resolved sandbox mode before relying on Docker settings. The source resolves agent-specific values ahead of defaults, with mode off, backend docker and workspace access none as fallbacks. That combination describes an available backend without enabling it. Inspect the agent override as well as the shared setting so you know which value applies to the session. Sandbox resolution.

When Docker sandboxing is selected, inspect its network, capability and mount configuration. The reviewed Docker defaults include network none and dropping all capabilities; the resolver also accepts configuration overrides. Those settings describe the selected container's intended restrictions. Workspace access and bind mounts determine which host data is passed into that environment. Docker defaults and bind configuration.

The official prompt-injection guide makes the off-mode consequence explicit: implicit host=auto resolves to the gateway host when no sandbox is active, while explicit host=sandbox refuses execution without an available sandbox runtime. Use that distinction when diagnosing why a task ran on the gateway or refused to start. The effective containment and escape resistance of a deployed backend require runtime verification beyond this static review. Sandbox and host selection.

Understand OpenClaw exec full, off and deny

Read all three exec fields together, then compare them with the effective policy output. The source constants supply fallbacks, and the resolver accepts overrides. The docs explain each field's job. Fallback constants; Resolution; Policy definitions.

Field and valueMeaning in practice
security: fullAn allowlist match is not required. It does not grant elevated access.
ask: offOrdinary policy prompts are disabled. Separately enabled strict inline-eval checks can still require approval.
askFallback: denyIf a prompt is required but no UI can answer, or it times out, block the request.
security: denyBlock host exec requests. This is different from denying only an unanswered prompt.

The practical consequence is that askFallback: deny does not turn full/off into an ask-first policy. A request allowed without an ordinary prompt never needs that unanswered-prompt fallback. Conversely, once a prompt is required, the fallback defines what happens if the approval channel is unavailable. Tool policy, host policy and applicable session modes still matter, so these exec findings do not establish universal send, purchase or deletion gates across integrations. Approval layers; Ask and fallback semantics.

For new configuration, the pinned docs prefer tools.exec.mode, with deny, allowlist, ask, auto and full modes. Full maps to the ordinary full/off behavior; ask uses allowlist policy and prompts on misses. Check the documented migration rules before combining this field with legacy security and ask keys: mixed policy can be rejected, and some legacy combinations have no exact normalized equivalent. Normalized exec modes.

Also inspect tools.exec.strictInlineEval if you allow interpreters. The docs describe approval requirements for recognized inline-evaluation forms and separate full-session/elevated exceptions. Use the relevant path's policy description when reviewing such commands, rather than assuming the word full overrides every other setting. Strict inline evaluation and exceptions.

Review OpenClaw scheduled work and tool availability

Check which gateway starts scheduled work and whether cron is enabled. The gateway constructs a CronService and resolves enablement from configuration. Trace an unattended workflow through that gateway and its target agent, then use the effective-policy checks for the host it can reach. A schedule identifies another entry point into the agent's permissions. Cron configuration; Cron service construction.

Inspect the enabled tools and plugins for that agent. The source builds tools for messaging, scheduling, web access, sessions and other operations conditionally, including plugins; an active count depends on configuration. The exec docs give a useful example: a messaging tool profile can exclude exec and process even when command approvals permit execution. Tool availability determines which action can be requested, approval policy determines whether it may proceed, and sandboxing determines its execution environment. Tool construction; Availability versus approval.

Review approval delivery for automation too. The pinned docs distinguish gateway-host cron approvals sent to capable operator clients from node-host automation's headless policy. Use the documented automation behavior for the chosen target before depending on a human to answer a scheduled request. Automation approval delivery.

Inspect OpenClaw memory, providers and diagnostics

Locate workspace MEMORY.md alongside the memory index when reviewing retention. Memory-core promotes records into that Markdown file and uses a SQLite-backed manager. Treat both as storage surfaces to inventory; deleting one file is not verified erasure of sidecar indexes, logs or recall history. Markdown promotion; SQLite-backed manager.

Check the configured provider as part of the same data review. The pinned default constants name openai and gpt-6-astra, but provider and plugin configuration can change destinations. Keep the model endpoint, messaging services and diagnostics exporter as separate entries in your connection inventory. Model and provider constants.

For the diagnostics OTel exporter, inspect diagnostics.enabled and diagnostics.otel.enabled. Its startup returns early when diagnostics configuration is missing, diagnostics is explicitly disabled, or OTel is not enabled. That establishes an opt-in condition for this exporter. Other telemetry defaults and an exhaustive effective outbound-host list are not established by these files. Exporter startup condition.

Use the comparison table to choose corresponding components in other projects. For example, the Hermes Agent review provides memory-file and terminal-backend checks, while the NanoClaw profile separates container mounts from provider approval prompts.

Check the scope of this OpenClaw security review

Use the reproduction record to distinguish the retained source slice from the attempted whole-tree scan. REA's full-tree retry reached the runner's 600-second cap without a complete JSON result. The retained graph analysis instead covers nine exact-byte cited-source files, with repository-relative paths and file hashes preserved. It parsed those files without parse failures; it is not a complete dependency analysis.

The earlier failed whole-tree attempt retained inventory evidence. A later runner metadata-release error required recovering timeout metadata from filesystem timestamps, not recovering a successful scan. The commands and output identities below preserve that distinction. Additional setup and security guidance in this page was read manually from the pinned repository documentation; it does not expand REA's nine-file coverage.

Reproduce this

Use the pinned revision and recorded commands below to trace the findings.

We read these with REA (morluto/rea, MIT, v6.4.0), an open-source reverse-engineering tool for agents, plus manual source reading. REA's repository supplies the tool; it does not endorse this review. We are not affiliated with this project or REA. This is static reading only: no target agent was executed and no live traffic was observed.

The pinned checkout is e46cb104f29782ab08caee36607946074f91056f. REA mapped the JS/TS application graph (Electron boundaries, imports, storage, request call sites); behavioural conclusions (approval paths, defaults, telemetry) were confirmed by reading the source. Electron boundaries were mapped where present. REA did not establish non-JavaScript core behavior.

These are portable equivalents of the commands recorded in the pass-2 manifest, run from a working directory that holds the pinned repository at repos/openclaw. Clone the repository at the pinned commit into repos/, create rea-out/, and run the commands from that directory. These are analysis commands, not instructions to start the agent. Executable and workspace paths are normalised to npx rea-agents@6.4.0 and relative paths; some original runs also set a writable compiler-cache directory through ADDITIONAL_SWIFT_DRIVER_FLAGS.

npx rea-agents@6.4.0 analyze-javascript-application repos/openclaw --format json --analysis-timeout-ms 120000 > rea-out/openclaw.json 2> rea-out/openclaw.stderr

npx rea-agents@6.4.0 analyze-javascript-application repos/openclaw --format json --analysis-timeout-ms 120000 > rea-out/openclaw-analysis.json 2> rea-out/openclaw-analysis.stderr

npx rea-agents@6.4.0 analyze-javascript-application rea-out/source-slices/openclaw --format json --analysis-timeout-ms 120000 > rea-out/openclaw-core.json 2> rea-out/openclaw-core.stderr

The retained analysis is partial. A writable compiler cache resolved the initial parsing setup failure, but blocked process-table inspection produced cleanup_incomplete. The runner stopped lingering CLIs after complete JSON output; that does not establish clean worker shutdown. Missing graph observations cannot establish absent behavior. The reproduction record records output hashes separately from Evidence IDs and subject digests.

For the bounded command, first recreate rea-out/source-slices/openclaw from the nine files listed in the reproduction record. Keep their repository-relative paths and verify their recorded SHA-256 values; do not substitute a reduced tree of your own choosing. The whole-tree timeout option applies per source, while the runner imposed the separate 600-second total cap.

FAQ

Is OpenClaw safe to use? It can be, on a machine and accounts you can afford to lose, after you change the defaults. Out of the box at 2026.10.2 the sandbox is off and host commands don't ask first. Keep the gateway on loopback, turn on sandboxing or approvals, install only skills you've read, and stay on a current release. The project had published 722 security advisories by 20 October 2026, a count that reflects how openly it discloses as well as how many bugs it fixed.

Why did OpenClaw delete all my emails? In the best-known case, from February 2026, the user's "don't action until I tell you to" instruction was lost when OpenClaw compacted a long context, and the deletions then ran as shell commands that needed no approval. An exec mode of ask or allowlist puts a gate outside the model, as long as the mail tool itself is not blanket-allowlisted, so a lost instruction no longer means unapproved deletions. The section above has the user's own posts.

How do I keep OpenClaw secure? Update to a current release, run openclaw security audit, keep the gateway off the open internet, choose an exec mode that asks or uses an allowlist, enable a sandbox for agents that read untrusted content, and install only skills and plugins you've reviewed.

Where should I start an OpenClaw security check? Run the official security audit, then inspect sandbox selection and the effective exec policy for the gateway or node that performs the work. The checklist above gives the documented commands.

Does askFallback deny make every command wait for approval? No. It blocks when a required prompt cannot be answered. Use the exec table above to distinguish that fallback from the policy deciding whether a prompt is required.

Is a missing exec tool the same as sandboxing? No. Tool policy determines whether an operation is available. The sandbox configuration determines the environment used for sandboxed execution. Inspect both before enabling a workflow.

How can I compare this with Hermes Agent? The Hermes Agent vs OpenClaw comparison sets the two projects' default execution, approvals and security models side by side. Use the Hermes Agent configuration checks, the ZeroClaw page (a Rust reimplementation built from OpenClaw's code) and comparison table to compare approval paths, execution environments and memory stores. Choose the deployment settings you need to inspect before comparing product labels.