Taiwan cybersecurity authority warns of OpenClaw AI agent risks
Taiwan's cybersecurity regulator warned in March 2026 that OpenClaw, a widely-adopted open-source AI agent tool, poses systemic security risks including exposure to malicious instructions, compromised third-party extensions, and degraded safety rules during extended operation. The authority recommended environment isolation, minimal privilege access, human review mechanisms, code inspection of extensions, and preservation of safety rules in persistent memory.
What happened
Taiwan's Digital Development Department Cybersecurity Administration issued warnings in March 2026 about security risks posed by OpenClaw, an open-source AI agent tool that has become widely adopted for automating daily tasks.
The authority identified three primary risk categories: malicious instructions embedded in external web content that AI agents might execute, malicious code hidden in third-party Skill extensions, and degradation of safety rules when AI agents compress memory during long-term operation.
Reported impact
The cybersecurity authority cautioned that without proper safeguards, AI agents like OpenClaw could serve as entry points for attackers to compromise personal devices and enterprise networks, potentially leading to data breaches, password theft, financial data exposure, identity theft and financial losses.
Index analysis
The authority highlighted that security risks are not attributable to single vulnerabilities but rather systemic architectural problems. Oasis Security researchers disclosed the ClawJacked vulnerability (CVE-2026-25253), which allowed attackers to conduct brute-force attacks on administrator permissions by luring users to malicious webpages without triggering browser warnings; the vulnerability was patched on January 29, 2026.
Governance controls
The cybersecurity authority recommended five protective measures: environment isolation using isolated computers, virtual machines or containers; minimizing external account privileges by creating dedicated accounts; implementing human review for high-risk operations; reviewing and scanning third-party Skill extensions before installation; and preserving safety rules in long-term memory files to prevent their loss during memory compression.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.