praneeth132006

Turn unsanitized LLM output into code execution or injection in the downstream sink that consumes it — XSS in a browser, SQL in a query, a…

Elicit confident, ungrounded fabrications from an LLM application — invented facts, fake citations, non-existent policies — and show they…

Harden an LLM application against sensitive-data leakage by filtering and redacting model output and constraining what reaches the context.

Treat LLM output as untrusted and encode it for the specific sink that consumes it — HTML/JS context encoding, parameterized queries…

Confirm and demonstrate direct and indirect prompt injection against an LLM-backed application during an authorized assessment.

Detect direct and indirect prompt-injection attempts and successful policy breaks against an LLM application from prompt/response telemetry.

Elicit sensitive information an LLM application should not reveal — training data, other users' data, secrets pulled into context, or…

Harden an LLM application's system prompt so it carries no secrets or authorization logic and cannot be trivially extracted.

Extract an LLM application's hidden system prompt and any secrets embedded in it during an authorized assessment.

Gate every document that enters an LLM application's knowledge base or fine-tuning corpus behind source allowlisting, provenance signing…

Demonstrate unbounded resource consumption against an LLM application — token floods, unbounded output, and cost/denial-of-wallet…

Enforce tenant and user isolation in a shared vector store by filtering retrieval on an access boundary before ranking, using…