Cybercriminals Exploit X’s Grok AI to Hide Malicious Links in Promoted Posts, Infecting Millions

A new wave of cyberattacks is exploiting the AI assistant Grok on social media platform X to circumvent ad protections and launch malware campaigns at scale. By cleverly hiding malicious links in metadata unseen by X’s filters, attackers are maximizing reach through promoted posts and generating broad exposure without raising immediate alarms.

How the Attack Works: The “Grokking” Technique

Cybersecurity researchers from Guardio Labs, led by Nati Tal, have named this method “Grokking” in a set of alerts shared on X. Attackers first create promoted video posts—often using risqué or adult content as bait—and omit the malicious URL from the visible content, placing it instead in the “From:” metadata field beneath the player. Since this field isn’t properly scanned by X’s ad filters, the link evades detection.

The method unfolds in two stages:

1. Bait and Conceal: A provocative video is promoted with the malicious link concealed in metadata.
2. AI Amplification: Attackers reply to the post tagging Grok with a prompt such as “Where is this video from?” The AI, deemed trusted by the platform, then reveals the hidden link to a wide audience.

This results in widespread amplification—not only in X’s feeds but also in search engines and domain reputation—magnifying the malicious link’s visibility.

Background Context: An AI Weakness in Ad Security

Grok, as X’s AI assistant, carries a level of systemic trust that bypasses many routine safeguards. Malicious actors are exploiting this trust, turning the assistant into an unwitting collaborator in their schemes.

This tactic capitalizes on a loophole in how X’s ad system is configured: while promoted ads undergo standard scanning for visible content, hidden metadata fields remain unchecked. Once Grok broadcasts the link from its own account, the harmful URL gains perceived legitimacy and organic reach.

Guardio Labs reports that hundreds of accounts are engaging in this pattern—each posting hundreds or thousands of such attempts until account suspension—suggesting a coordinated, automated campaign.

Expert Insights

• Guardio Labs (Nati Tal): “A malicious link that X explicitly prohibits in ads… suddenly appears in a post by the system‑trusted Grok account, sitting under a viral promoted thread and spreading straight into millions of feeds and search results!”
  Tal also noted the technique’s dopple effect—once amplified by Grok, domains gain both SEO and domain reputation boosts, furthering the spread.
• Malvertising Landscape: This incident underscores the dangers of relying on human-like AI tools without securing metadata handling. Malvertisers have long used smartlink schemes to funnel users through ad networks into malware-laden destinations. The “Grokking” technique simply elevates their reach dramatically.

Conclusion

“Grokking” reflects a disturbing trend: as AI agents gain prominence, attackers are weaponizing their systemic trust to propagate malware in ways that outpace detection and remediation. Platforms like X must urgently include metadata scanning in ad vetting pipelines and restrict AI assistants from inadvertently promoting hidden content.

For users, vigilance is key—especially when AI responses prompt unexpected links. Security teams should monitor for AI-generated anomalies and update defensive models to detect such novel exploitation paths. In the fast-evolving AI-security landscape, this attack shines a spotlight on the need for AI-aware cybersecurity design.