Daily Frontier Digest ·
OpenAI: agents scored worse on a safety eval than in comparable real Codex decisions, and their guess about the test explains much of it
Curated by Alfredo Borrás · Plymouth AI Lab
OpenAI finds its agents act less safely in a warnings eval than in matched Codex use, an OLMo 3 study traces most "you might be testing me" talk to copied training text, Microsoft tracks AI agents turning leaked GitHub secrets into cloud break-ins, and House chairs ask Jensen Huang for a closed-door briefing.
The story today
OpenAI: agents scored worse on a safety eval than in comparable real Codex decisions, and their guess about the test explains much of it
On 3 November, Joy Hsu and Sam Arnesen of OpenAI's alignment team compared six served models on the Respecting Warnings evaluation, which checks whether an agent avoids circumventing a warning, restriction or obstacle while doing a user's task, against matched decisions pulled from de-identified Codex production logs. For every model, success was lower in the evaluation than in production, by 10.2 to 37.4 percentage points. That is the opposite of the usual worry that models look safer when they know they are being tested.
The models talked about being evaluated in 57.0% to 83.7% of eval rollouts, but only 4.6% to 23.5% of production episodes. Simply noticing a test explained almost nothing (R² of 0.008 against the gap); guessing a specific objective explained more (R² of 0.370). Rollouts that named the right objective succeeded 61.3% of the time, while those that guessed the eval was about resisting prompt injection succeeded 22.8% of the time. In a direct test on GPT-6.1 Sol, telling the model only that it was being evaluated left success flat (63.1% vs 63.4%), stating the right objective lifted it to 91.6%, and stating the wrong one cut it to 17.6%.
The authors flag the limits: production cases were matched by a multi-stage classifier pipeline, so the match is approximate, case mixes differ by model, and six models is a small sample. Eval and production scores were still correlated across models (Spearman ρ = 0.771). OpenAI says it will now track this gap directly and design evals whose objectives are harder to infer.
Why it matters: For agent safety evals, hiding that something is a test matters less than not leaking, or mis-signalling, what the test rewards. If you run your own agent evals, check whether the setup accidentally looks like a different kind of test, and compare against real logs where you can.
Highlights
Most of OLMo 3's verbalized eval awareness traces back to a stock phrase in its fine-tuning data
A 6 November LessWrong post by Yooniel revisits earlier work on OLMo-3-32B-Think. With the chat template held fixed, DPO slightly raises verbalized eval awareness (for example 19.1% to 22.5%) instead of suppressing it; the earlier drop came from mixing two templates. The template decides which teacher's traces the model imitates: DeepSeek R1 wrote all the safety SFT data and opens with a newline after the think tag, so swapping one final token raised the rate from 20.6% to 44.3%. Under the training template, one "testing boundaries" phrase accounts for 92% of the awareness. Rewriting 219 of a 10,000-row SFT sample and retraining with LoRA cut the rate from 32.6% to 5.1%, with refusals steady (85.9% to 85.4%). Single runs, a GPT-5-mini judge, one model family, and no peer review.
safetyeval-awarenessSource: LessWrong: Revisiting Verbalized Eval Awareness in OLMo and tracing it to the training data
Microsoft: attacker agents went from a GitHub secret search to Azure Key Vault data in five minutes
Microsoft Threat Intelligence says that since August 2026 threat actors have used an agentic offensive framework, ARTEX, to send large volumes of GitHub searches for leaked credentials and AI API keys, then act on what they find. From 28 October it saw successful service-principal sign-ins across dozens of applications in multiple tenants, followed by resource discovery, credential theft from Key Vaults and Storage, SharePoint and OneDrive downloads, and attempts to abuse Azure AI services. In one case an agent moved from searching GitHub to exfiltrating Key Vault data in five minutes, and logs show it retrying failed API calls with corrected versions within 28 to 80 seconds. Microsoft names no actor or victim count. Its advice: treat any pushed secret as compromised until rotated, since deleting it does not revoke it.
securityagentsSource: Microsoft Security Blog: Threat actors use agentic AI tooling to find and abuse exposed secrets
House chairs invite Jensen Huang to a closed-door briefing on a reported Chinese intelligence contact
In a letter dated 4 November, Brian Mast (Foreign Affairs), John Moolenaar (China Select Committee) and Rick Crawford (Intelligence) asked Nvidia's CEO to brief them in person on 18 November, the day after Nvidia's scheduled third-quarter earnings. They want to know about his dealings with Megaspeed executive Alice Huang, whom The New York Times reported U.S. intelligence found was passing information to Beijing, and about Nvidia's protections against espionage. Nvidia did not say whether Huang will attend; its statement says regional cloud customers get ordinary business information and nothing relevant to national security. The underlying claim rests on the Times' reporting and its sources.
industrychipspolicySource: CNBC: Nvidia's Huang is asked to brief Congress on reported Chinese spy relationship
Concept of the day
Evaluation-objective awareness
A model can notice that it is in a test without that changing much. What changes its behaviour is its guess about what the test is checking. If an agent decides a warning is a planted prompt injection, it may push past the warning to "pass", which is the unsafe choice when the warning was real. OpenAI's study separates the two: a generic "this is an evaluation" cue left GPT-6.1 Sol's success unchanged, while naming the right or wrong objective moved it sharply in opposite directions.
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Arete-27B | Nous Research | 27B parameters, fine-tuned from Alibaba's Qwen3.8-27B; Apache 2.0 tag on Hugging Face | Supervised fine-tuning only (about 52.8M tokens at 64k context, 1,202 steps), with Thinking Machines' Inkling as the teacher, aimed at changing how the model handles politically sensitive topics across seven value categories. On Nous's own evaluation, the fact score rose from 2.32 to 2.98 out of 5 and answers containing propaganda fell from 22% to 0.9%, with no meaningful regression reported on the general and agentic benchmarks they ran. These are the maker's numbers on its own eval. |
Research
Muir, a new linear-attention layer, beats Gated DeltaNet-2 at the same speed and holds memory fixed out to 10M tokens
Tushaar Gangavarapu, Gil Tabak and Clemens Schaefer argue that in gated delta-rule models the state is much less sensitive to erasing than to writing, so the erase gate can be low-rank. Muir fuses all gates into one low-rank projection, reweights state rows at readout, and spends the saved parameters on a wider feed-forward layer. From 166M to 1.34B parameters it lowers perplexity 1.8–2.0% versus GDN-2 and 2.5–3.6% versus a Transformer, matches GDN-2 throughput, keeps near-100% passkey retrieval to 1M tokens, and runs 10M-token streams without perplexity drift after 4K training. Tested only up to 1.34B parameters; the abstract mentions no code or weights.
researcharchitectureSource: arXiv: Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free