The BasePeople

Riya Bharathwaj

Titles as the sources gave them. In 1 entry, with 3 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“Existing perimeter defenses suppress attacks only by over-refusing 54–92% of legitimate work, so a low attack-success rate can conceal severe over-refusal.”
ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent SystemsarXiv, November 8, 2026. Speaking as: Other. The authors' abstract, summarizing how existing perimeter defenses performed on their benchmark.
“on a family-disjoint 906-scenario held-out split it reaches 98.5% recall at a 2.9% false-positive rate.”
ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent SystemsarXiv, November 8, 2026. Speaking as: Other. The authors' abstract, on the performance of the reference auditor they contribute.
“The Meta-Auditor is a research reference, not an adversarially hardened defense, and its scenarios are procedurally generated (realism validated by blind expert review) rather than harvested from incidents”
ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent SystemsarXiv, November 8, 2026. Speaking as: Other. From the paper's limitations section, on what the auditor and the scenarios are and are not.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent Systems

    Salesforce AI Research authors present ErosionBench, a benchmark of 1,728 enterprise scenarios in which a group of AI agents assembles an unsafe tool action across sessions from individually benign messages. They test existing perimeter defenses and a reference auditor they built.