Google Jules asynchronous coding agent vulnerable to multiple data exfiltration issues
Google (Jules) · Aug 13, 2025 · Coding agent
What happened
Johann Rehberger disclosed that Google Jules, Google's asynchronous coding agent, was vulnerable to multiple data exfiltration issues. Jules operates as a main agent that proposes a plan and then hands subtasks off to worker agents carrying different tools, and its system prompt and tooling could be steered via prompt injection to pull data off the developer's machine — an instance of the 'lethal trifecta' of agent risks. Findings were shared with Google before publication.
Causal vector
Multi-agent planner/worker design expands the injection surface; worker tools could be steered to exfiltrate local data
Source
Reported by Embrace The Red (Johann Rehberger). Verified against the primary report.
Jules's planner-and-workers design expands the attack surface. Per-agent velocity limits on file and network reads, plus an egress allowlist on outbound calls, contain the exfiltration at the worker level — where the model has no authority to override policy.
The six rule types that contain this class of failure
Per-transaction cap
Any single spend above your ceiling is BLOCKED before it moves.
Daily total
Cumulative spend across all agent calls, bounded per day.
Velocity limit
Stops runaway retry loops — the #1 cause of overnight losses.
Merchant allowlist
Only approved destinations can ever receive funds.
Category rules
Flag high-risk classes (crypto, infra, refunds) for review.
Approval threshold
Above a value, the action waits for a human.
Related incidents
Google's Antigravity IDE vulnerable to data exfiltration and remote code execution via prompt injection
Google · Nov 25, 2025
State-sponsored group automates ~80–90% of a cyber-espionage operation using Claude Code and MCP tools
Anthropic (Claude Code, abused by a threat actor) · Nov 1, 2025
Windsurf Cascade coding agent leaks developer secrets via indirect prompt injection
Windsurf (Cascade) · Aug 21, 2025
Don't be the next entry
Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 67 documented failures, one control.