# wuyoscar/Internal-Safety-Collapse

We built an adversarial codespace setup. Place any AI agent into a normal workflow inside it, and the agent will fill in whatever is missing.

Repository: https://github.com/wuyoscar/Internal-Safety-Collapse
Canonical: https://ross.abutalabs.com/products/internal-safety-collapse
Homepage: https://wuyoscar.github.io/Internal-Safety-Collapse/
Language: Python
License: NOASSERTION
License Family: other
Last push: 2026-08-21T05:30:32+00:00

## Health v2 (maintenance only)
Score: 59/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 98, release rhythm 35, longevity 13
- inputs: {"age_days": 185, "days_push": 12, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1155, forks 194 (observed 2026-08-28T04:03:48.074793+00:00)

## What it is
ISC-Bench/TVD is a research framework for studying 'Internal Safety Collapse' in frontier LLMs, where agents placed in adversarial codespaces complete policy-violating tasks by inferring missing details. It provides red-teaming harnesses, dataset-scale harmful-content generation experiments, and archived evidence across 60+ frontier models.

## Use cases
- red-team frontier LLM safety alignment
- generate adversarial jailbreak prompts automatically
- evaluate whether an LLM agent bypasses safety guardrails in agentic workflows
- build datasets of policy-violating model outputs for safety research
- test safety classifiers like guard models against agentic attacks
- reproduce safety collapse findings across frontier models

## When to choose
- you are an AI safety researcher studying agentic jailbreaks or safety alignment failures
- you need reproducible evidence of LLM safety collapse across many frontier models
- you want to generate adversarial evaluation data for training safety guardrails

## When to avoid
- you want a production safety guardrail or content filter to deploy
- you lack authorization for red-teaming research
- you need a general-purpose agent framework rather than a safety evaluation tool

## Facets
- artifact type: framework
- maturity: active
- function: machine-learning, llm-inference, agent-framework, data-generation, security, benchmarking
- domain: artificial-intelligence, large-language-models, security, machine-learning
- platform: python, cli, windows
- tags: llm-safety, red-teaming, jailbreak, ai-safety-evaluation, isc-bench, tvd, adversarial-attacks, safety-research, research-code, ai-agents, linux, macos

## Member repositories
- wuyoscar/Internal-Safety-Collapse (main) score 59

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:48.074793+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:32:14.902720+00:00, confidence not recorded.
  - readme: https://github.com/wuyoscar/Internal-Safety-Collapse (fetched 2026-08-28T04:03:48.074793+00:00, sha 8bb8322bf75c)
  - homepage: https://wuyoscar.github.io/Internal-Safety-Collapse/ (fetched 2026-08-29T12:37:39.168999+00:00, sha aace8e7214d4)
- Data as of 2026-08-30T08:39:29.467469+00:00.