Please (Don't) Stop the Music: Adversarial Red-Teaming of AI...
summary
Red teams two music-generation services with a corpus of 275 banned German songs, finds roughly 70–75% bypass of default filters producing about ten hours of harmful audio, and demonstrates simple bypasses (phonetic edits, hyphenation, encodings). Builds an output-check pipeline: extract vocals, transcribe with Whisper → GPT‑4 classifier into ten harm categories and JSON, then pass/fail; this reduces successful harmful outputs to about 4.4% (≈45 minutes) but notes transcription errors, stitchable verses, and prompt tricks remain.
AI-written summary · about
chapters (10)
- 0:00AI alignment and risks
- 1:17Outer vs inner alignment
- 3:26Policy enforcement in AI
- 4:52Germany's music laws
- 6:39Red-teaming music models
- 8:49Bypassing input filters
- 9:57Detecting harmful lyrics
- 11:46Limitations of VR3 encoding
- 12:52AI safety takeaways
- 13:56Call for AI policy
related talks
50mFace the Music: Cybersecurity and the Music Industry
24mCommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition
37mHow to Break AI Systems (Before Someone Else Does)
50mResponsible AI: Adversarial Attacks on LLMs
30mHidden Voice Commands
32mSecurity Implications of Artificial Intelligence Synthesizers
53mHow to Break AI Systems (Before Someone Else Does)
43mTalk by Om Thakkar (OpenAI)