Agentic AI tools like OpenClaw (formerly ClawdBot / MoltBot) introduce a new category of security risk. Because these agents run persistently with access to files, email, APIs and online services, they present a much larger attack surface than traditional software.
Key Threat Categories (as of Feb 2026)
Threat
Description
Infostealer / Credential theft
Malware (e.g. Vidar variants) scans config directories for keywords like “token” and “private key”, exfiltrating gateway tokens and API keys from files like openclaw.json
Prompt injection
Malicious instructions hidden in web pages, emails or documents hijack the agent’s behaviour, causing it to exfiltrate credentials without triggering conventional alerts
Remote code execution (RCE)
Hundreds of thousands of exposed OpenClaw instances have been found, creating pivot points for attackers who can execute arbitrary code via a single exposed service
Malicious third-party skills
Bad actors upload poisoned skills to ClawHub, sometimes bypassing VirusTotal by hosting payloads on lookalike sites rather than embedding them in SKILL.md files
Memory poisoning
Adversarial instructions planted in an agent’s long-term memory persist across sessions, causing it to take harmful actions days or weeks after the initial compromise
Covers end-to-end AI system security, adversarial attacks, and AI-specific threat models — directly maps to prompt injection and agentic attack surfaces
Focuses on real-world breach case studies, NLP-based attack techniques, and mitigating attacks on generative AI models — covers the credential-theft angle
Paid (LinkedIn Learning)
Course
Why Relevant
Search: “AI security” or “prompt injection”
LinkedIn Learning’s catalogue in this area is thinner than Coursera’s; check for updated 2025/2026 courses on AI agent security as the catalogue is growing quickly