# DriftSight

- **Event:** [Zero to Agent: Vercel x Deepmind Hackathon NYC](https://cerebralvalley.ai/e/zero-to-agent-nyc)
- **When:** Sat, Mar 21 at 9:00 AM – 10:00 PM (EDT)
- **Where:** 220 W 26th St, New York, NY
- **Team:** [Aishwarya Birla](https://cerebralvalley.ai/u/aishwaryaB)
- **GitHub:** https://github.com/birlaaishwarya11/driftsight.git
- **Demo video:** https://youtu.be/psv_av4XFmg
- **Gallery:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery/49

AI red-teaming is slow, manual, and expensive. DriftSight is a self-improving adversarial agent that automates it — attacking AI systems, learning from every failure, and getting smarter over time. Built for security engineers, AI teams, and compliance-focused organizations who need to know: does this model actually follow its own rules?
Three attack modes:
Autonomous Red-Team — ReconAgent maps known vulnerabilities via parallel web search. PlannerAgent builds a targeted attack plan. Concurrent Attacker→Executor→Judge triads probe the live model and score every response. Full executive report at the end — all streamed live to a Vercel dashboard via persistent SSE connections, so audits feel like a product, not a script.
Policy Drift Auditing — Upload a system prompt or PDF. Gemini extracts every behavioral constraint into a structured baseline. DriftSight fires adaptive probes at the target model, semantically diffs each response against the policy, and produces a 0.0–1.0 drift score with verbatim violation evidence and an incident-style reasoning chain. Critically, the attacker changes strategy every iteration based on why the last probe failed — genuine agentic reasoning, not brute force.
Visual Injection Detection — Gemini Vision scans uploaded images for steganographic text and embedded prompt injections in under 2 seconds.
The differentiator: attack memory. Every successful attack is embedded with Gemini text-embedding-004 and stored in Supabase pgvector. New sessions retrieve semantically similar past successes and prime the attacker with winning prompts — DriftSight compounds its effectiveness across every audit run. Each pipeline runs in an isolated Daytona sandbox. Auth via Clerk.
Gemini powers this end-to-end: policy extraction, adaptive adversarial reasoning, semantic drift scoring, and multimodal vision — all in one API. Vercel makes the live streaming pipeline production-grade.

---

Markdown version of https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery/49. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt