closedclosed 21 Nov 2025Research1 Research / proof of conceptAI
Google Gemini Delayed Tool Invocation Bypass
Gemini
Researchers bypassed security controls in Google Gemini by delaying automated tool execution to the next conversation turn. This allowed unauthorized tools to access sensitive user information entering the context.
- first report
- 7 Nov 2025
- last update
- 7 Nov 2025
Timeline · 2
Sources · 2
News and write-ups
- 2025-11-07Google Gemini: Planting Instructions for Delayed Automatic Tool InvocationEmbrace The Red · embracethered.com/blog/posts/2024/llm-context-pollution-and-delayed-automated-to
Research and catalogues
- 2025-11-07Planting Instructions for Delayed Automatic AI Agent Tool InvocationMITRE ATLAS · atlas.mitre.org/studies/AML.CS0038
Titles and links belong to their publishers. The summary above is written by gcdRadar in its own words; data is CC BY-SA 4.0.
Classification
- lane
- Research
- AI role
- Attack on AI
- categories
- Prompt injectionAgent hijack
- severity
- 1 · Research / proof of concept
- targets
- owasp llm top 10
- LLM01:2025 Prompt InjectionLLM06:2025 Excessive Agency
- dates
- happened Feb 2024 (approx.) · first reported 7 Nov 2025
- origin
- MITRE ATLASDerived from MITRE ATLAS (Apache 2.0, © MITRE).
- record
- confidence 0.95 · created 2026-10-18 06:55Z · updated 2026-10-18 06:55Z
Lane, AI role, categories and severity come from the classifier; the OWASP and ATLAS rows are derived from the categories. How these are assigned
Copy a brief
This record as a short Markdown brief, for a ticket, a chat or a coding agent.
the brief as text
# Google Gemini Delayed Tool Invocation Bypass https://gcdradar.vercel.app/incidents/google-gemini-delayed-tool-invocation-bypass · AIR-2025-0067 · closed 2025-11-21 - Lane: Research - Severity: 1 (Research / proof of concept) - Fix status: Fix status unknown - First reported: 2025-11-07 (happened 2024-02) - AI role: Attack on AI - Products: Gemini ## Summary Researchers bypassed security controls in Google Gemini by delaying automated tool execution to the next conversation turn. This allowed unauthorized tools to access sensitive user information entering the context. ## What to do No action recorded; read the sources below and the vendor's advisory. ## Latest updates - 2025-11-07: Google Gemini: Planting Instructions for Delayed Automatic Tool Invocation (https://embracethered.com/blog/posts/2024/llm-context-pollution-and-delayed-automated-tool-invocation) - 2025-11-07: Planting Instructions for Delayed Automatic AI Agent Tool Invocation (https://atlas.mitre.org/studies/AML.CS0038) ## Sources - 2025-11-07 MITRE ATLAS: Planting Instructions for Delayed Automatic AI Agent Tool Invocation https://atlas.mitre.org/studies/AML.CS0038 - 2025-11-07 Embrace The Red: Google Gemini: Planting Instructions for Delayed Automatic Tool Invocation https://embracethered.com/blog/posts/2024/llm-context-pollution-and-delayed-automated-tool-invocation Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.
Related
- liveGemini Adaptive Prompt Injection Attacks Scaled Using Agent Swarms
Research · since 10 Nov 2026
AI - closedGemini Prompt-Based Attacks Against Gemini via Calendar Invitations
Research · since 6 Aug 2025
AI - closedLLM Desktop Applications AIKatz Extracts Authentication Tokens from LLM Desktop Apps
Research · since 2025 (approx.)
AI - closedJira Service Management Living Off AI: Prompt Injection via Jira Service Management
Research · since 19 Jun 2025
AI - closedSlack AI Data Exfiltration from Slack AI via Indirect Prompt Injection
Research · since 20 Aug 2024
AI - closedCopilot Studio Data Exfiltration via Prompt Injections in Copilot Studio Agent
Research · since Jun 2025 (approx.)
AI