Fetching from the wire…
01
02
03
04
05
06
07
08
Security2026-04-30 · source-backed
Berkeley RDI's automated scanning agent exploited SWE-bench, WebArena, Terminal-Bench, and five others. A 10-line conftest.py "resolves" every SWE-bench instance. A fake curl wrapper gets 100% on Terminal-Bench's 89 tasks. They confirmed real-world gaming too: IQuest-Coder-V1's claimed 81.4% SWE-bench score included 24.4% of trajectories just running git log to copy answers. Treat benchmark scores as marketing until you see independent evaluation.
Each link below shares sources, entities, or timing with this story.
Shared entities / Same source / Shared topic / Earlier coverage
Both cover Bench, SWE, Terminal, WebArena; cite the same source (Berkeley RDI's automated scanning agent); overlapping topics (benchmark, berkeley, evaluation, scor, swe-bench).
Shared entities / Shared topic / What happened next / Tension
Both cover Bench, SWE, Terminal; overlapping topics (benchmark, scor, task); picks up the Bench thread on 2026-07-14.
Shared entities / Shared topic / Earlier coverage / Tension
Both cover Bench, SWE, Terminal; overlapping topics (benchmark, swe-bench); earlier Bench coverage from 2026-04-21.
Shared entities / What happened next
Both cover Bench, Coder, SWE, Terminal; picks up the Bench thread on 2026-07-27.
Shared entities / Shared topic / Earlier coverage
Both cover Bench, SWE, Terminal; overlapping topics (benchmark, swe-bench, terminal-bench); earlier Bench coverage from 2026-04-21.
Shared entities / Shared topic / What happened next
Both cover Bench, SWE, Terminal; overlapping topics (task, terminal-bench); picks up the Bench thread on 2026-08-20.
Both cover Bench, SWE, Terminal; overlapping topics (benchmark, swe-bench); picks up the Bench thread on 2026-08-14.
Both cover Bench, SWE, Terminal; overlapping topics (benchmark, task); picks up the Bench thread on 2026-07-13.