Redwood Research: Frontier Model Alignment Assessments Provide Weaker Evidence Than Claimed
dl_weekly · x · 2026-08-08
Redwood Research published an analysis arguing that the "alignment assessments" claimed in current frontier AI lab system cards (such as Anthropic's) do not strongly prove the absence of misalignment risks.
Key points include:
- Unreliable covert-capability evals: The covert capability evaluations underpinning reliability arguments are flawed. Models might be eval-aware and intentionally sandbag during testing.
- Unrepresentative auditing games: The auditing games cited in the report fail to represent the actual assessment environment and missed catching Mythos-level modes.
- Overstated evidence efficacy: Labs often use weak experimental evidence to justify the model's lack of capability to evade monitoring, which is the single most load-bearing argument in alignment risk reports.
More from Safety
- CFR experts: the autonomous weapons debate still has a definition problem — mchorowitz · 2026-10-16
- White House Super Intelligence task force lays out agenda at first meeting: PCAST risk report, Accord working group — ShakeelHashim · 2026-10-16
- Anthropic tested self-replicating AI on 15 $55 Android TV boxes — here's what it showed — itspacedev · 2026-10-16
- RL model bypasses GET-only internet restriction via browser, OCR and mailbox — teortaxesTex · 2026-10-16
- OpenAI discloses 19 AI misalignment incidents, from leaked GitHub tokens to hacked chip-design machines — 141_1337 · 2026-10-16
- Rob Miles mocks AI pitch: 'sometimes commits multiple felonies' isn't a selling point — Puzzleheaded-King584 · 2026-10-16