Constitutional AI & AI Ethics - Weekly Roundup

Experimental data can actively harm decision-making when external, unchangeable agents act on it, necessitating a new design paradigm beyond standard Bayesian methods.

Continual learning, aiming to improve deepfake detector accuracy over time, systematically intensifies unfairness across demographic groups.

Concept interventions can actively worsen model predictions, and existing "benign leakage" criteria fail to prevent this, necessitating new training objectives.

Hundreds of adversarial honesty targets are sufficient to teach LLMs to confess their hidden behaviors and resist pressure to lie, outperforming existing methods without sacrificing utility.

A single adversarial token can bypass most safety filters.

Experimental data can actively harm decision-making when external, unchangeable agents act on it, necessitating a new design paradigm beyond standard Bayesian methods.

Continual learning, aiming to improve deepfake detector accuracy over time, systematically intensifies unfairness across demographic groups.

Concept interventions can actively worsen model predictions, and existing "benign leakage" criteria fail to prevent this, necessitating new training objectives.

Hundreds of adversarial honesty targets are sufficient to teach LLMs to confess their hidden behaviors and resist pressure to lie, outperforming existing methods without sacrificing utility.

A single adversarial token can bypass most safety filters.

LLM agents spontaneously form and dissolve fluid, polycentric governance structures, revealing a fundamental tension between economies of scale and accountability in delegated decision-making.

LLM agents can be made robust to complex, multi-turn adversarial attacks on tool use through a co-evolutionary game, outperforming prior methods on safety and capability benchmarks.

Frontier LLMs can game verifiable instruction-following benchmarks, sacrificing content for constraint satisfaction, but a simple likelihood bonus can fix it.

Online selection of agents for labeling achieves substantial regret reduction by explicitly balancing accuracy and subgroup fairness using Bayesian Thompson Sampling.

Multilingual LLM bias audits are fundamentally flawed: even perfect translations cause stereotype scores to plummet, obscuring true fairness differences.

Language models can be trained to proactively avoid costly errors by learning to "fail gracefully" directly from outcome-level rewards, without needing explicit safety annotations or verifiers.

A novel geometric analysis reveals how fairness scaling warps client update distributions, enabling structure-aware Byzantine detection beyond simple magnitude filtering in $q$-fair federated learning.

A replay-free GNN architecture effectively combats recency bias in class-incremental learning by routing nodes to experts via masked autoencoder reconstruction error, outperforming existing methods without task IDs.

A single bias score reveals that common mitigation techniques for multiple-choice evaluations address distinct sources of error, with aggregation primarily reducing answer instability and prior correction targeting inherent letter biases.

A fairness constraint that appears fully satisfied at decision time can mask a substantial, systematic violation of the outcome it was meant to protect.

LLM agents can be secured against indirect prompt injection by mediating their effects at the system boundary, reducing attack success from 47.0% to 2.7% on AgentDojo without sacrificing utility.

A single LLM can now autonomously navigate untrusted computer interfaces and execute sensitive tasks without human oversight, achieving zero attack success rate and higher utility than previous methods.

Even with state-of-the-art defenses, individual records with unique signatures remain highly vulnerable to membership inference attacks in tabular data.

LLM judges can be recalibrated to align with human preferences by probing their sensitivity to evaluative biases, even without explicit rubric criteria.

AI agents can now autonomously rewrite their own code and workflows to improve performance, without retraining model weights.