Three equal orange columns followed by a descending row of blue columns

Every Rate on This Leaderboard Is a Whole Number of Tax Returns

Surge AI’s DAYJOB: Tax leaderboard publishes pass rates to one decimal place. Every one divides into whole trials out of 250, and second and third place are one tax assignment apart.

October 29, 2026 · 6 min · 1227 words
A balance scale with arrows pointing both ways beside a one-directional ramp

The Same Overlap, Measured Two Ways

Project Zero compared a fuzzer against an LLM bug-finding agent on Chrome’s V8. Three filters produce two changes of lead, and the two overlap figures measure different things.

October 28, 2026 · 7 min · 1327 words
Two horizontal bars, one short and one long with a lit segment, showing two different time windows

The Exposure Window Appears in Only One of the Two Reports

Cloudflare and the SGLang maintainers published coordinated reports on the same cross-request KV cache leak. Both are candid; the blast radius is in one and the twenty-month exposure window is in the other.

October 27, 2026 · 6 min · 1260 words
A two-bar chart, a tall off-white bar at 190 beside a rust bar at 95

Cerebras Published the Arithmetic. One Assumption Halves It

Cerebras published every step of fitting a 3T-parameter model onto 59 wafer-scale chips, ending at 190 concurrent requests. The architecture checks out; one disclosed cache assumption, which its own docs contradict, halves the answer.

October 26, 2026 · 6 min · 1125 words
Five glass vials in a row, the first four bracketed together and the fifth glowing green

The Variable the Press Release Never Named

A press release and its own preprint agree on every number. The paper’s only surviving predictor falls outside the sentence the release was making.

October 23, 2026 · 6 min · 1135 words
A lit platform where a path reaches a columned building, above a dim one where a rust bar stops the same path

WhatsApp's Support AI Was Told Not to Recommend a Regulator

WhatsApp’s hidden support-AI instructions told the chatbot not to recommend regulators or cite privacy laws in cases where its public policy names those routes.

October 20, 2026 · 4 min · 765 words
A magenta ribbon resolving into three token tiles, with a faint second path diverging at one junction

HybridAgent: Readable Is Not Yet Faithful

HybridAgent makes latent reasoning readable, but its evidence tests relevance and readability rather than semantic faithfulness.

October 19, 2026 · 4 min · 776 words
A dot plot of time-to-first-token across four hardware configurations, one point clearly worse than the rest

Apple's Best MLX Number Lives in a Tab a Fetcher Never Gets

Apple’s MLX post verifies on every load-bearing figure. Its best number sits in a table tab that is not server-rendered, so a fetcher-based reader never sees it. A full recount of all 32 parallelism cells.

October 16, 2026 · 5 min · 977 words
A horizontal bar chart of vulnerability categories, two short bars among several long ones

Three Fixes Work, Two Do Not, and SSRF Stays at 2.4%

An agent benchmark where the top performer on functional tests wrote the least secure code. Three interventions measurably help, two do not, and two weakness families stay near zero throughout.

October 15, 2026 · 7 min · 1480 words
Four plugs sparking on contact, one plug connecting cleanly

Four Elicitation Methods Failed, the Fifth Guessed All 256 Triggers

Five elicitation methods were run against a model with a chemistry classifier encrypted into its weights. Four returned chance accuracy. The fifth recovered it by enumerating all 256 possible triggers, which the authors say themselves.

October 14, 2026 · 7 min · 1297 words