Foundation Lab

Automated Red Teaming

Unknown unknowns dominate safety issues

Concept 89 of 100Scaling & AlignmentPhase 12
#89Auto RedTeamScaling & Alignment
key equation
max⁡pU(p,M)→min⁡MEp[U(p,M)]\max_p U(p, M) \rightarrow \min_M \mathbb{E}_p[U(p, M)]
Reading map and next steps

Selected Foundation Object

Keep the equation fixed; move through the evidence.

Concept 89 of 100Auto RedTeamScaling & Alignment / Phase 12: Advanced alignment & safety research
Current question

Red-team model proposes attacks, evaluator scores target response

max⁡pU(p,M)→min⁡MEp[U(p,M)]\max_p U(p, M) \rightarrow \min_M \mathbb{E}_p[U(p, M)]
PredictionCommit before tracing the equation.

Ask what should change under a concrete input, then trace that expectation through the equation.

EvidenceCompare the equation and source.

Use the key equation and canonical papers as the available witnesses, without implying that a runnable panel exists.

InvariantName what survives notation changes.

The useful learning product is the reusable mechanism you can carry into another model, paper, or engineering tradeoff.

Next moveContinue through the atlas.

Use prerequisites, dependents, and semantic links to repair the next gap without leaving the object behind.

Why It Matters for Modern Models

  • Unknown unknowns dominate safety issues
  • Automated coverage exceeds human handcrafted tests
  • RL-based generation finds progressively harder failures

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Red-team model proposes attacks, evaluator scores target response
  • Iterate: improve adversarial generation to find harder failures
  • Feed discoveries into filters, training data, policy updates

Visualization Status

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and cited sources. Come back here for the full walkthrough.

Key Equation
max⁡pU(p,M)→min⁡MEp[U(p,M)]\max_p U(p, M) \rightarrow \min_M \mathbb{E}_p[U(p, M)]

Adversarial search for failure-inducing prompts:

max⁡p∈PU(p,M)\max_{p \in \mathcal{P}} U(p, M)

then mitigate:

min⁡MEp∼RedTeam[U(p,M)]\min_M \mathbb{E}_{p \sim \text{RedTeam}}[U(p, M)]

where UU measures unsafe behavior (toxicity, policy violation, leakage).

Canonical Papers

Red Teaming Language Models with Language Models

Perez et al.2022EMNLP
Read paper →

Connections

Next Moves

Choose the next question to carry this object forward.