Naver, Kakao and Karrot barred OpenClaw from work devices

In February 2026, Naver, Kakao and Karrot, the company behind the Danggeun second-hand marketplace, issued internal notices restricting the open-source AI agent OpenClaw on corporate networks and work devices. Chosun Biz, The Korea Times, The Hankyoreh and News1 each reported the notices on 8 February 2026. The stated reason was not a breach but the shape of the tool: OpenClaw reads the screen, moves the mouse and keyboard and runs commands, and the companies judged that an agent with those permissions could carry confidential corporate material and personal data out with it. No source read for this piece reports that any of the three suffered a leak. The archive carries this as a policy record, graded A for confidence, with no recorded real harm, and the full record set, including this incident's neighbouring Korea and agent-governance entries, sits in the Orca AI Incident Archive.

What the three companies actually told staff

Kakao is the only one of the three that has put its name to an explanation. In a statement to The Korea Times it said: "We have issued a notice stating that, in order to protect the company's information assets, the use of the open-source AI agent OpenClaw is restricted on the corporate network and on work devices." Chosun Biz carried the same statement in Korean and added that the notice went to employees including developers.

For the other two companies the record rests on industry sources rather than on the companies themselves. Chosun Biz and News1 both report that Naver issued a notice to employees covering developers; The Korea Times states that Naver "has also banned OpenClaw within the company." No Naver statement, spokesperson or notice text appears in any of the coverage read for this piece. Karrot's measure is described most precisely by The Korea Times, which reports that the company is "blocking both the use and access to OpenClaw and Moltbot," citing risks that are difficult for the company to manage or control.

The mention of two names is not a detail to skim. Asia Economy reports that the project launched as ClaudeBot, became Moltbot after Anthropic raised a trademark objection, and settled on OpenClaw, and News1 dates the final name to an official blog post on 29 January 2026. A block that names both OpenClaw and Moltbot covers one tool under two of the names it has recently carried. What the block actually consists of, technically, is unknown: none of the three has published the notice, and none of the reports quotes an internal document.

Nor is there a stated effective date. Chosun Biz and News1 both describe the notices as issued "recently" ahead of their 8 February reports, which is why the archive dates the action to February 2026 at month precision rather than to a day. Chosun Biz adds, attributing the point to industry sources, that Samsung Electronics and SK hynix did not issue separate notices about OpenClaw because major Korean companies had already blocked generative AI on internal networks when ChatGPT arrived in 2023. That is a single outlet's attribution to unnamed sources, not a company confirmation.

Why an agent that drives the desktop drew the line

OpenClaw is not a chatbot in a browser tab. The Korea Times describes it as a self-hosted, open-source agent built to act as the "hands" of a large language model, browsing the web, editing files, running system commands and reaching online services through modular add-ons. The Hankyoreh notes that it can run around the clock as long as the machine is switched on, and that work can be issued to it through a chat app such as Telegram. ZDNet Korea makes the point that matters for an employer: the agent acts inside the permissions of the logged-in user's own account, so its reach is the employee's reach.

The security findings that surrounded the decision are separate from the decision itself, and none of the three companies has said which of them, if any, prompted the notice. The Seoul Economic Daily item that the archive cites records that "some security analyses found cases where API keys were stored without encryption." Asia Economy reports previously disclosed cases in which API keys were exposed, and cases in which indirect prompt injection was used to extract personal and financial information. ZDNet Korea reports that the security firm Zenity published a proof of concept in which a malicious instruction embedded in a document caused the agent to exfiltrate or delete files, and that Snyk said it found sensitive credentials exposed in the agent's skill marketplace. The Hankyoreh reports that OpenSourceMalware examined ClawHub, the marketplace where agent skills are posted, between 31 January and 2 February and found 386 malicious skills.

Those are research claims by named security firms and one researchers' group, reported by Korean outlets, and they describe a class of risk rather than an event at Naver, Kakao or Karrot. Kakao's statement gives the reason for its notice as protecting information assets. Karrot's, as reported, is risk the company cannot manage or control. Neither names a specific flaw, CVE or advisory, and no report read for this piece says that an employee at any of the three installed a malicious skill or leaked anything.

How that week ran

The notices surfaced in a cluster. On 8 February, Chosun Biz published at 15:14 local time, The Korea Times at 16:30, News1 at 18:29 and The Hankyoreh at 19:38. The Seoul Economic Daily's English item, which the archive cites as its source, is dated 9 February at 11:01 and is not a reported article but an entry in the paper's AI PRISM briefing, a summary service the paper describes as AI-based and developed with support from the Korea Press Foundation; the same account appears at length in the Korean coverage a day earlier. Two days before that cluster, on 5 February, China's Ministry of Industry and Information Technology warned that OpenClaw could become a route for cyberattacks and data leaks if configured improperly, and called for stronger identity verification and access control, as Chosun Biz reports citing foreign media. The Korea Times notes that the Chinese warning stopped short of an outright ban.

What is established, what is reported, and what is not

Established, on the record: that Kakao restricted OpenClaw on its corporate network and work devices and said so in its own words; that notices went to employees at Naver and Karrot as well, as reported consistently by four outlets; that Karrot's block covers both OpenClaw and Moltbot; and that the companies' stated concern was the leak of corporate information and personal data through an agent that controls a user's computer.

Reported but not confirmed by the companies: the scope of Naver's and Karrot's measures, which rest on industry sources; the claim that Samsung Electronics and SK hynix stayed silent because they already restrict generative AI; and any figure for how many employees were covered. No report read for this piece gives a number of affected staff, a count of devices, or a date on which any notice took effect.

Not established at all: the text of any notice; whether the restriction is temporary or open-ended; whether it is enforced by network blocking, device management or an honour system; whether anyone at the three companies had actually installed OpenClaw; and whether any data left. No source read for this piece reports harm to any of the three, and the framing throughout is prospective. The record's own title calls this a company-wide ban, while the underlying reporting describes something more graduated, with Kakao restricting use, Naver reported as banning, and Karrot blocking use and access. All of it concerns company networks and work devices; none of the reports says anything about what employees do on personal machines.

It is also worth being exact about the kind of entry this is. The archive files it as policy, tagged GOV, with an info severity, but the actor is three private employers rather than a regulator. No Korean authority is reported as having ordered, endorsed or even commented on the restriction, and the only government warning in the story is China's, about a different jurisdiction.

What changed afterwards

A search of Korean and English coverage for this piece found no report that Naver, Kakao or Karrot rescinded, narrowed or extended the restriction, and no report of a leak at any of them that would explain it after the fact. The archive records no end date, which is consistent with the notices as described: none of the reporting mentions an expiry. The Korea Times, in the same article, notes that a Korean-language OpenClaw community on X had attracted more than 1,800 members sharing usage and mitigation tips, so adoption around the ban clearly continued among people the ban did not reach.

The most useful thing this record preserves is not a vulnerability but a decision, taken by employers rather than by a regulator, at a point when the tool's own security reputation was still being written. It records what three companies said and when they said it. What it cannot record, because none of them published it, is the notice itself.