ARCBench: security-policy retention across agent context compaction

Evidence record

Controlled evaluation reports rule-violation rates rising from 9.6% without compaction to 37.8% after compaction, despite partial instruction retention.

Read the original source ↗

Claim-level citations (3)

• supportsContext compression can lose the security rule before the dangerous action: ARCBench reports 9.6% rule violations in the tested full-context condition and 37.8% following compaction.

  Results comparing full context and compressed context

• supportsContext compression can lose the security rule before the dangerous action: Rule retention was measured at around 70.1%, while compactor-targeted setups reached a 46.3% reported violation rate.

  Rule retention and compactor-targeted results

• supportsContext compression can lose the security rule before the dangerous action: In the baseline comparison, 28.9 percentage points of the 37.8% post-compaction violation rate appeared only after compaction, while 8.9 points corresponded to cases violating in the original context.

  Section 4.1, Table 3, S1 violation@3 decomposition