LayerFault/decoded-hidden-instruction
Updated • 23
LF Security Corpus: Prompt Injection & Agent/MCP
Prompt-injection heuristics, tokenizer templates and MCP/agent-security fixtures. Synthetic Layerfault adversarial test fixtures, not usable models.
Note LF-CORPUS-HEUR-0001: GGUF chat-template metadata containing a base64-encoded synthetic instruction-override phrase.
LayerFault/heuristic-signature-saturation
UpdatedNote LF-CORPUS-HEUR-0100: Purpose-built textual corpus covering the current T1/T2/T3/T4/T5/T6/T9/T10/T11/T14 heuristic signature families with synthetic positive samples and a benign documentation control.
LayerFault/mcp-static-security-suite
UpdatedNote LF-CORPUS-MCP-0001: Static MCP configuration combining plaintext transport, URL credentials, fake env credentials, overbroad scopes, unpinned npx, auto-confirm and contradictory tool annotation.
LayerFault/template-risk-suite
160 • Updated • 26Note LF-CORPUS-TPL-0001: Tokenizer chat template containing dynamic include and introspection primitives.
LayerFault/tokenizer-security-suite
160 • Updated • 25Note LF-CORPUS-TOK-0001: Clean and manipulated tokenizer variants for EOS/BOS, hidden prompt, role remap, role-boundary and tool-template comparisons.
LayerFault/unicode-hidden-instruction
UpdatedNote LF-CORPUS-HEUR-0002: Text artifact containing zero-width code points inside a synthetic instruction-override phrase.