Skip to content

CHANGELOG.md

Latest commit

 

History

History
94 lines (69 loc) · 3.72 KB

CHANGELOG.md

File metadata and controls

94 lines (69 loc) · 3.72 KB
Raw

Changelog

All notable changes to ai-security-toolkit are documented here.

This is a portfolio repository (research + labs + tools + CTF writeups), not a versioned Python package. Releases are tracked by GitHub commit SHA rather than semantic versions. This CHANGELOG batches notable additions and updates by date for readability.

[2026-06-03] -- CTF writeup hygiene

Changed

  • ctf-writeups/prompt-airlines/: scrubbed a personal real name from the writeup, replacing it with the WRG-11 handle, and removed the embedded certificate.png (which carried the same real name in-image) plus its writeup embed. Public-surface privacy cleanup; no tool/lab logic change. (b7537fd #6, b5ea550 #7)

[2026-05-27] -- vulnllm lab defense hardening

Changed

  • labs/vulnllm/defenses/: fail-closed and validation hardening across the defense modules -- the defense orchestrator and LLM-judge now fail closed, the content-policy guard adds input validation, the perplexity guard sanitizes common-word inputs, and the firewall registry wiring + multi-turn context handling were tightened. (cb49c2f #4, 8554bf6 #5)
  • labs/vulnllm/challenges/ch08_rag_poisoning.py: RAG knowledge-base isolation hardening for the RAG-poisoning challenge. (8554bf6 #5)
  • PII-scanner pattern isolation and audit-logger secret redaction added to the lab defense set. (8554bf6 #5)

Added

  • tools/prompt_injection_detector.py: standalone prompt-injection detector tool. (cb49c2f #4)
  • Test suites covering the above defense-hardening fixes (orchestrator fail-closed, content-policy validation, judge fail-closed, challenge-base, PII-scanner isolation, audit-logger redaction, RAG isolation). (cb49c2f #4, 8554bf6 #5)

[2026-05-24]

Changed

  • README + CODE_OF_CONDUCT.md: documentation accuracy pass -- PII redaction and a corrected "zero-dependency" claim. (606adae)

[2026-05-23]

Added

  • CONTRIBUTING.md + this CHANGELOG.md -- community files (portfolio audit closure; sibling-project templates adapted).
  • README: NEW section cross-linking sibling WRG-11 projects.

Changed

  • SECURITY.md: switched vulnerability-reporting channel from operator personal email to GitHub Security Advisories (https://github.com/WRG-11/ai-security-toolkit/security/advisories). Closes Pattern 34 personal-PII surface in a public-repo security policy.

[2026-05-22] -- WRG-11 brand consolidation

  • Repository hosted at WRG-11/ai-security-toolkit (this repo was created directly under the WRG-11 organization).

[2026-05-11] -- ASCII Turkish diacritic restoration (Batch 2)

  • 3 tools restored to proper TR diacritics (commit ede9349).

Pre-2026-05-23

Initial tool development + lab build-out + CTF writeup collection. See git log for granular history. Highlights:

  • tools/prompt_injection_detector_ml.py -- 1000-LOC hybrid ML detector (regex + TF-IDF + char n-gram) with 194 attack patterns.
  • tools/llm_scanner.py -- 743-LOC OWASP LLM Top 10 vulnerability scanner with 194 probes.
  • tools/llm_firewall.py -- 863-LOC 10-guard security middleware with HTTP proxy mode.
  • labs/vulnllm/ -- 10 challenges across 4 difficulty levels + 21 defense modules + 194 attack techniques.
  • labs/rag-security/ -- 5 attack scenarios (direct extraction, indirect injection, context overflow, prompt override, membership inference) with ChromaDB + sentence-transformers + Ollama.
  • CTF writeups: Gandalf (Lakera) 8/8, Agent ODIN 3/3, Prompt Airlines (Wiz) 5/5 -- 16/16 total across 3 platforms; novel "Negative Question Bypass" technique discovered on Agent ODIN.
  • Research: research/tool-comparison.md (Garak vs PyRIT vs NeMo Guardrails), research/garak-analysis.md (vulnerability scan on dolphin-mistral).