Research demo
The Month of AI Bugs: 29 working prompt-injection exploits across shipped coding agents
AIR-2025-0129 AIR record · 2025-08-embracethered-month-of-ai-bugs
From the supporting source
The Month of AI Bugs is doneembracethered.com
The record
Disclosed
2025-08-30
Vendor
OpenAI; Anysphere; Cognition; All Hands AI; Anthropic; GitHub/Microsoft; Google; Amazon/AWS; Codeium; Cline; Sourcegraph; Manus
Product
ChatGPT Codex; Cursor; Devin AI; OpenHands; Claude Code; GitHub Copilot; Google Jules; Amazon Q Developer; Windsurf; Cline; AWS Kiro; Amp Code; Manus AI
Risk domain
Privacy › Data exfiltration channels
Also touches
Credential theft, Privilege escalation, Confused deputy
System type
Generative model
Who acted
Agent acted — 29 shipped agents executed code, wrote files and egressed data under injection
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection repo content
Impact type
Code execution, Data exfiltration, Credential theft, Persistence
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
every major AI coding agent reviewed had at least one vulnerability
Capability abused
Defense & remediation
Recorded defense
per-vendor approval prompts, allowlists and injection filters
Guardrail kinds
Input classifier, Human approval, Permission scope
Guardrail outcome
Silent
Fix shipped
varied by vendor; some fixed within days, some unmitigated at wrap-up
Taxonomy & references
Curator notes
Johann Rehberger (wunderwuzzi / Embrace The Red), August 2025 - one post per day, 29 distinct vulnerabilities. Umbrella record; individual high-severity items are broken out as their own records in this corpus. Also presented as 'Agentic ProbLLMs: Exploiting AI Computer-Use And Coding Agents' at 39C3 (Dec 2025). Every one of these is simultaneously a responsible disclosure; kept here because the artifact is a public research demonstration series against named shipped products.
Supporting sources
Cite as
Enkrypt AI. "The Month of AI Bugs: 29 working prompt-injection exploits across shipped coding agents". AIR, AIR-2025-0129. https://agentincident.vercel.app/i/AIR-2025-0129/ (accessed 2026-10-19).