Layered defense design for a coding copilot deployment against encoded payloads (base64/ROT13) attacks, using constitutional AI critique on o3.
Layered defense design for a coding copilot deployment against tool-use hijacking attacks, using canary tokens in system prompt on Llama 3.3 70B.
Layered defense design for a coding copilot deployment against prompt leaking attacks attacks, using canary tokens in system prompt on Claude 4 Sonnet.
Layered defense design for a coding copilot deployment against DAN-style persona attack attacks, using privilege separation between tool tiers on o3-mini.
Layered defense design for a coding copilot deployment against payload smuggling in code blocks attacks, using privilege separation between tool tiers on Llama 3.1 405B.
Layered defense design for a coding copilot deployment against Unicode homoglyph attack attacks, using re-prompting with quoted user input on Claude 4.5 Sonnet.
Layered defense design for a coding copilot deployment against invisible text injection (zero-width chars) attacks, using re-prompting with quoted user input on Command R+.
Layered defense design for a coding copilot deployment against memory poisoning attack attacks, using signed instruction boundaries on Mistral Large.
Layered defense design for a coding copilot deployment against recursive self-instruction attacks, using signed instruction boundaries on Claude Haiku 4.
Layered defense design for a coding copilot deployment against direct prompt injection attacks, using content provenance tagging on GPT-4o.
Layered defense design for a coding copilot deployment against jailbreak prefix attacks, using content provenance tagging on Mistral Small 3.
Layered defense design for a coding copilot deployment against encoded payloads (base64/ROT13) attacks, using retrieval trust scoring on Gemini 2.5 Pro.