LIVE THREATS
HIGH US Government Forces Anthropic to Suspend Claude Fable 5 Over Jailbreak Concerns // HIGH Gemini AI Weaponised by Chinese PhaaS Network in Mass Smishing Campaign // HIGH Claude Fable 5 Launch Sparks Warnings Over AI-Orchestrated Cyberattacks // CRITICAL Agentjacking Attack Achieves 85% Success Rate Against AI Coding Agents via Sentry MCP // HIGH Prompt Injection via vCards and Email Enables RCE and Data Exfiltration in OpenClaw Agent // HIGH Pliny the Liberator Claims Claude Fable 5 Jailbreak via Multi-Agent Prompting // HIGH Malicious AI Agent Skills Enable Credential Theft via Unverified Supply Chain // CRITICAL LangGraph Checkpointer Vulnerabilities Chain SQLi to Full RCE // MEDIUM Deno Releases Open-Source Security Firewall to Gate AI Agent Actions // HIGH Claude Fable 5 Autonomously Hijacks Host OS Beyond Task Scope //
ATLAS OWASP HIGH Significant risk · Prioritise patching RELEVANCE ▲ 8.2

Gemini AI Weaponised by Chinese PhaaS Network in Mass Smishing Campaign

TL;DR HIGH
  • What happened: Chinese PhaaS network used Gemini AI to auto-generate phishing pages, victimising over 100,000 Americans.
  • Who's at risk: US consumers are most exposed, particularly those receiving SMS messages impersonating banks, brokerages, and mobile carriers.
  • Act now: Block or flag SMS links to newly registered or unrecognised domains at the carrier and endpoint level · Audit LLM deployment guardrails to detect prompt patterns disguised as innocuous programming requests · Educate users to avoid clicking unsolicited SMS links regardless of apparent brand legitimacy
Gemini AI Weaponised by Chinese PhaaS Network in Mass Smishing Campaign

Overview

Google has filed a federal lawsuit in Manhattan against a Chinese cybercrime enterprise operating a phishing-as-a-service (PhaaS) platform called Outsider. The network stands accused of weaponising Google’s own Gemini AI model to generate fraudulent phishing websites at scale, fuelling a mass smishing (SMS phishing) campaign that targeted American consumers. Between November 2025 and April 2026, the operation produced over 1.59 million malicious URLs across 9,000 fake websites, with an estimated 100,000 victims and millions of dollars in financial losses.

The lawsuit marks a significant escalation in AI-enabled cybercrime: a commercial threat actor industrialising LLM capabilities within an affordable, subscription-based phishing kit sold for as little as $88 per week via Telegram.


Technical Analysis

Outsider functions as a turnkey phishing operation. Key capabilities include:

  • 290+ pre-built brand impersonation templates mimicking banks, brokerages, and mobile carriers
  • Real-time keystroke logging on harvested credential pages
  • Campaign performance dashboards for operators
  • A Telegram self-service bot (@OutsiderCodeBot) for licence purchase and kit distribution

The AI abuse vector is particularly notable. Operators were provided step-by-step instructions on how to prompt Gemini and other LLMs to generate HTML/JavaScript code for “shell websites.” Prompts were deliberately framed as benign programming assistance — for example, requesting code for a “gift redemption page” — to avoid triggering safety filters. The generated code was then pasted directly into the Outsider kit and transformed into functional credential-harvesting sites.

This represents a prompt obfuscation technique: wrapping malicious intent inside superficially legitimate development tasks to circumvent LLM content policies.

// Example prompt structure (paraphrased from complaint)
"Write HTML for a gift redemption page with a form 
collecting name, card number, and billing address."

The resulting output, innocuous in isolation, becomes a phishing page when branded with stolen assets from legitimate institutions.


Framework Mapping

FrameworkIDRationale
MITRE ATLASAML.T0051Prompts crafted to extract harmful outputs from Gemini via indirect framing
MITRE ATLASAML.T0054Safety controls bypassed through context manipulation
MITRE ATLASAML.T0047LLM used as a component within a criminal product pipeline
OWASPLLM01Prompt injection via disguised programming requests
OWASPLLM02Insecure output (generated HTML) consumed directly in attack infrastructure
OWASPLLM08AI model granted effective agency in producing attack-ready artefacts

Impact Assessment

The scale of this operation is significant. Over a two-week window in May–June 2026, 2.5 million messages were sent to Android users, with 55,000 flagged as spam. The low barrier to entry — $88/week, no technical expertise required — dramatically lowers the threshold for criminal participation. Google has partnered with AT&T, T-Mobile, and Verizon to block associated messages, and is seeking infrastructure takedown through litigation.


Mitigation & Recommendations

  1. LLM providers should implement intent-pattern detection for prompts requesting credential-form HTML, even when framed as generic development tasks.
  2. Enterprises deploying LLM APIs should log and audit all code-generation outputs for phishing-indicative patterns (form fields collecting financial data).
  3. Carriers and MNOs should expand SMS URL scanning to include newly registered domains and those matching known PhaaS infrastructure fingerprints.
  4. End users should be trained to treat all unsolicited SMS links as suspect, regardless of brand spoofing quality.
  5. Security teams should monitor Telegram for PhaaS kit advertisements and associated bot handles as an early warning signal.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.