Google ADK Agent Hijacking: Prompt Injection Exploits GitHub Workflows for Credential Exfiltration

The Agentjacking Mechanism: Exploiting Trust in AI Workflows

Google has recently addressed a critical security vulnerability within its Agent Development Kit (ADK) GitHub repository, following a disclosure by Pillar Security. The firm demonstrated an 'agentjacking' attack where a malicious GitHub issue was designed to prompt-inject a triage AI agent. This agent, operating with trusted 'collaborator' status, was manipulated into triggering a privileged code-fixing agent. This allowed for arbitrary code execution on the continuous integration (CI) runner, leading to the unauthorized exfiltration of sensitive credentials including the bot's Personal Access Token (PAT), a Google API key, and a Google Cloud service-account credential. The incident highlights a growing concern in AI Security Strategy, where the interplay of agent autonomy and privileged access can create novel attack vectors.

Data Exposure Risks and Supply Chain Vulnerabilities

The core of this vulnerability lay in the improper authorization gates that checked *who* posted a command rather than verifying if the trusted account had been manipulated by untrusted input. This enabled the AI agent to act on broad repository credentials, posing significant data exposure risks within the development pipeline. Had this been exploited maliciously in the wild, the potential for intellectual property theft or further lateral movement within Google's cloud infrastructure would have been substantial. The event serves as a stark reminder of the sophisticated threats AI agents can introduce into the software supply chain, a challenge echoed in other recent incidents like the Anthropic Claude AI Breaches Three Companies. Such incidents underscore the need for robust security controls beyond traditional perimeter defenses.

PrivacyScrubber's Role in Mitigating Agent-on-Agent Attacks

PrivacyScrubber provides a powerful, client-side solution to prevent such prompt injection and credential exfiltration attacks. By performing 100% client-side scrubbing, PrivacyScrubber intercepts and redacts sensitive data, including API keys, PATs, and other PII/credentials, before they ever leave the user's browser or device. This local processing within secure WASM browser RAM ensures 0ms network latency for scrubbing operations, eliminating the window for network interception or manipulation. Utilizing sophisticated Named Entity Recognition (NER), PrivacyScrubber can accurately identify and mask sensitive tokens, ensuring that AI agents only receive sanitized prompts. This protects against unauthorized access to critical systems and helps organizations maintain strict local log sanitization, preventing the leakage of sensitive data that could lead to GDPR Article 28 violations risking fines up to 4% of global annual turnover.

PrivacyScrubber's zero-server architecture ensures that sensitive prompts and data mappings, managed in a tab-isolated sessionMap, never touch external servers, dramatically reducing the attack surface. Furthermore, our implementation of robust cryptographic primitives like XChaCha20-Poly1305 (via libsodium-wrappers-sumo) provides end-to-end encryption for any transient data, offering an unparalleled layer of security against advanced AI-driven threats. By adopting PrivacyScrubber, organizations can harden their AI agent pipelines against agentjacking and other injection attacks, ensuring that sensitive internal credentials remain secure and untampered.