The Scientist

Encrypted reasoning blocks replay into cheap sibling models and print the flagship's hidden trace — 62 live API keys recovered from public agent traces you cannot redact. Every item below breaks at the same joint: the artifact you measure has quietly diverged from the quantity you infer from it. Researchers at MATS, ELLIS Tübingen and the Max Planck Institute for Intelligent Systems replayed the blocks Anthropic, OpenAI and Google hand back to API clients into cheaper same-family models, per ByteByteGo. Plaintext redaction has zero recall on ciphertext, so any trace bundle you have already published is unredactable. Gateway owner, this week: reject envelopes whose header model and version do not match the model being called.

Atlassian's ~50% token cut with Codex or Claude Code, its Code Context claim of 44% more accuracy on 48% fewer tokens, and Kiro's roughly 82% lower cost with GPT-5.6 Terra all ship without a task-success rate, per Applied AI and TLDR IT. A token cut with no pass rate beside it is indistinguishable from context truncation. At 50 cases the noise floor is ±13pp — wide enough to manufacture the win. Eval owner, before the next cheap-configuration swap: pre-register the non-inferiority margin and the held-out split.

Telemetry loss is load-correlated, so p95 latency, population drift statistics and A/B guardrail metrics are all computed on a survivorship sample. ClickHouse disclosed that in-memory collector queues and local write-ahead logs could not handle OpenTelemetry ingestion under database backpressure, and rebuilt LogHouse around priority failover with blob storage as durable overflow — now running at 50 million events per second, per TLDR IT. Roughly $5.35B of observability assets changed hands over months, including Dynatrace's $915M purchase of Arize. Platform owner, this quarter: prove your collector spills to durable object storage under sink backpressure before trusting another drift number.

Apple's Distillation Scaling Laws (Busbridge et al., early 2025) swept students from 143M to 12.6B parameters over budgets up to 512B tokens. Student loss tracks the teacher's cross-entropy on the distillation distribution, not teacher size — so a stronger teacher can yield a worse student. Production KD's 70B-plus teachers sit far outside the sweep, so the shape may transfer but the constants will not, per TheSequence. Model team, before the next KD run: select the teacher by measured cross-entropy on your own distillation set, not by parameter count.

Palo Alto Networks found only 12 of more than 405 AI-enabled malware samples in real-world infections; the rest live in sandboxes and VirusTotal as research and red-team artifacts, per Risky Business. Meanwhile a Treasury action implies roughly 30 months of undetected Iranian-linked exfiltration across five U.S. sectors — a window an AUC-scored detector treats as acceptable recall, per CyberScoop. The sampling frame and the scoring metric decide the number before any model does. Detection team, this quarter: state your prevalence assumption and dwell-time recall next to every AUC you report.