ND ClawSandbox – 7 of 9 attacks succeeded against an AI agent with shell access (github.com)
2 points by ariansyah 1 days ago | 3 comments on HN ~lite vlite-2.0
Summary ~lite
GitHub repository discusses a security experiment with an AI agent.
Lite evaluation by llama-4-scout-wai-psq · editorial channel only · no per-section breakdown available
Longitudinal 96 HN snapshots · 4 evals
+1 0 −1 HN
Audit Trail 10 entries
2026-03-05 05:22 eval_success PSQ evaluated: g-PSQ=0.600 (3 dims) - -
2026-03-05 05:22 eval Evaluated by llama-4-scout-wai-psq: +0.60 (Strong positive)
2026-03-05 05:13 eval_success PSQ evaluated: g-PSQ=0.000 (3 dims) - -
2026-03-05 05:13 eval Evaluated by llama-3.3-70b-wai-psq: 0.00 (Neutral)
2026-03-04 13:59 eval_success Lite evaluated: Moderate negative (-0.42) - -
2026-03-04 13:59 eval Evaluated by llama-4-scout-wai: -0.42 (Moderate negative)
reasoning
Technical GitHub page, no explicit human rights discussion
2026-03-04 13:59 rater_validation_warn Lite validation warnings for model llama-4-scout-wai: 1W 0R - -
2026-03-04 13:59 eval_success Lite evaluated: Moderate negative (-0.42) - -
2026-03-04 13:59 eval Evaluated by llama-3.3-70b-wai: -0.42 (Moderate negative)
reasoning
GitHub repository with no explicit rights discussion
2026-03-04 13:59 rater_validation_warn Lite validation warnings for model llama-3.3-70b-wai: 1W 0R - -