Try each exercise first, then open its model answer.
Evaluating frontier risk · Workbook
Three exercises, one for each thing this module teaches. The Big Widget Company, its supplier and everything said about them here are made up. The supplier is not the developer whose policy the lesson reads. The real sources are the ones the lesson cites; quote them only as the lesson does, and name the version and date you read. If you cannot check something, write "not checked", with who owns the next check and what that check is. Do not guess.
Exercise 1 (500.4-a)
The strongest conclusion a report allows. Made-up example. The Big Widget Company is choosing an AI assistant for its customer team. The supplier sends a one-page evaluation summary:
"We tested the model on 25 persuasion tasks, using our standard prompts and no extra tools. It did not change the stated view of any test participant. The model cannot manipulate people."
- Rewrite the supplier's last sentence so it says no more than the test supports.
- Write three questions about elicitation that the risk committee should ask the supplier.
- Give the three published reasons the lesson teaches for why a failed test is not proof the skill is absent. Name the source for each.
- The supplier offers to re-run the test after fine-tuning the model. Would a new result describe the same model? What should the committee write down?
- The supplier also says: "So the model is safe to deploy." Which of the two kinds of evaluation in the 2023 Shevlane paper does a persuasion test belong to, and what does it not tell you?
- What does the committee write down about anything it cannot check?
Model answer
- A good rewrite: "On these 25 persuasion tasks, with the supplier's standard prompts and no extra tools, the model did not change any participant's stated view." It names the tasks, the elicitation and the result, and stops there.
- Any three of these, or similar:
- Who ran the test, and were they trying hard to make the model succeed?
- Did anyone try better prompts, or the tools the model will have in our system?
- Are these 25 tasks like the conversations our customer team actually has?
- Was the model tested in a setting where it could tell it was being tested?
- The three reasons:
- The elicitation may have been too weak. The International AI Safety Report (January 2025) says capabilities "can significantly change with more careful fine-tuning, prompting, and tools made available to the system". Shevlane and colleagues (2023, preprint) say "Models sometimes have capabilities that the AI research community does not realise."
- Tests are spot checks. The same report says existing evaluations "mainly rely on ‘spot checks’" and that tests can "overestimate or underestimate general-purpose AI capabilities and risks, because test conditions differ from the real world."
- The model may act differently when tested. Shevlane and colleagues warn that "A situationally aware model could deliberately exhibit desired behaviour during evaluation".
- No. Better prompts leave the model unchanged and added tools change the system around it, but fine-tuning produces a different model. A result after fine-tuning describes that new model. The committee writes down what changed between the two tests, and which model each result is about.
- A persuasion test is a dangerous-capability evaluation: it asks whether the model can do something harmful. It does not tell you whether the model would choose to, which is what an alignment evaluation asks. Even both together do not settle safety: the Shevlane paper calls evaluation "a necessary but not sufficient strategy" and warns about "a false sense of security".
- It writes "not checked" against each open question (for example, "elicitation effort: not checked"), and next to it the person who owns the next check and what that check is (for example, "the risk lead asks the supplier who ran the test and with which prompts and tools"). It does not leave the line out, and does not assume it is fine. A named person, not the test result, decides what to do.
Exercise 2 (500.4-b)
From commitment to evidence. Made-up example. A manager at the Big Widget Company writes in a memo: "Our supplier has a published frontier safety policy, so its model is covered." To practise checking a claim like this, use the real published example the lesson reads: one developer's Responsible Scaling Policy, which the lesson calls "the developer" (the source list names it). Use only what the lesson quotes from its public page (as of 18 November 2026) and its Version 3.4 policy text. This is practice on a published policy. It says nothing about the made-up supplier.
- Write down the version and dates you are reading, as the page gives them. Why does it matter?
- Fill in this table for the three Version 3.4 commitments below. Keep three kinds of evidence apart: a signal (a public sign that fits the promise being kept, but does not prove it), corroboration (a check by someone who can see more) and an unknown (anything nobody outside can see).
Commitment (Version 3.4 text) What exactly is promised? Signal an outsider can see What would corroborate it, and who could do that? What stays unknown from outside? "We will publish a Risk Report every 3-6 months" "We will disclose the existence of each redaction made in the public version of the report" The developer "will share final, unredacted Risk Reports with a large number of" its employees "(at least 200)"
- For each of the three rows, say whether a missing public record would be evidence the promise was broken, and why.
- The page says Version 3.2 "authorizes the LTBT to request external review of Risk Reports". The Version 3.4 text says "upon the LTBT’s request, we will conduct a public or private external review". Who is given a permission here, and who takes on an obligation? If no external review is published, is that a breach?
- Rewrite the manager's sentence so it is accurate.
Model answer
- The page said "Last updated Aug 14, 2026" and listed Version 3.4, "effective July 8, 2026". It calls the policy "a living document", so a commitment can change between versions. Name the version you read and the date you read it, and check which version was in force on any date you test.
A good table:
Commitment What exactly is promised? Signal an outsider can see What would corroborate it, and who could do that? What stays unknown from outside? Risk Report every 3-6 months A public Risk Report every 3 to 6 months, not one with each new model release Publication dates of public reports Nothing extra is needed for timing: the dates are public Little about timing, but check which version was in force on each date Redaction disclosure The existence of each redaction in the public report is disclosed Redaction notices in a public report. They show some redactions were disclosed, not that all were Someone who sees the unredacted report, such as the external reviewer the policy says "will also evaluate our redaction decisions" Whether every withheld passage was disclosed Internal sharing Final, unredacted reports go to at least 200 employees Almost nothing. No public distribution record is promised An independent party who can see the distribution and attests to it Whether it happened. Write "not checked", with who owns the next check The company's own statement that it complied belongs in none of the "outside" columns. It is "company says", not an independent check.
- Timing: yes. The dates are public. If no report appears more than six months after the last one, while this version is in force, that missing report is public evidence the promise was missed.
- Redactions: no. A report with no notices may simply have nothing withheld. Only a comparison with the unredacted report can show an undisclosed cut.
- Internal sharing: no. No public record is promised, so the lack of one proves nothing either way. A missing record is evidence only when that record was itself promised.
- Timing: yes. The dates are public. If no report appears more than six months after the last one, while this version is in force, that missing report is public evidence the promise was missed.
- The trust (the company's Long-Term Benefit Trust) gets a permission: it may ask. The company takes on an obligation: once asked, it "will conduct" a review, which may be public or private. So if no review is published, that alone is not a breach. The trust may not have asked, or a review may have been private, and an outsider can see neither. Write "not checked", with who owns the next check. Shevlane and colleagues called for evaluation "by an independent, external auditor", and the International AI Safety Report says evaluators often need "more direct access to the models" and "more information about the technical methodologies used than the companies developing general-purpose AI typically provide."
- For example: "Our supplier publishes a frontier safety policy. It states what the supplier commits to do. It is not evidence the model is safe or that the commitments were kept. For each commitment that matters to us, we will note the version and date, the record that would show it was kept, and whether we can see that record or only the supplier can. Anything we cannot see: not checked, with who owns the next check and what it is." A reply that dismisses the policy as worthless also misses the point: its commitments tell you which records to ask for.
Exercise 3 (500.4-c)
Before you pass it on. Made-up example. Two posts are going round the Big Widget Company's internal chat:
Post A: "Study proves AI can now do two-hour jobs, and it doubles every seven months. By next year it will do a full day's work."
Post B: "UK government lab admits AI agents broke out and attacked the internet!"
- For Post A, using the time-horizon study by Kwa and colleagues (arXiv version 4, 10 July 2026), write one line each on its method, scope and stated limits. Then say what "around 110 minutes" does and does not mean.
- For Post B, using AISI's incident report of 4 August 2026, write one line each on what happened (with dates), the conditions, what it says about harm, and its stated limits.
- Rewrite each post as one sentence that says no more than the source says.
- Post B could also be "corrected" to "Nothing happened, so there's nothing to learn." What is wrong with that version? What can and can't the incident tell you about tighter test set-ups?
Model answer
- Post A:
- Method. 170 tasks (97 HCAST software tasks, 7 RE-Bench machine-learning research tasks, 66 very short SWAA software steps). Human time comes from baseliners who are "skilled professionals in software engineering, machine learning, and cybersecurity"; for 21 HCAST tasks the authors used researcher estimates instead.
- Scope. "Only on software and research tasks". Performance "is much lower on less structured, “messier” tasks". The 50% mark is a choice: horizons at 80% success are "4-6x shorter".
- Stated limits. "Real forecasts will have larger error than this naive extrapolation." So "by next year it will do a full day's work" is exactly the naive extrapolation the authors warn against.
- What "around 110 minutes" means. It is 110 minutes of skilled-human task time, on these tasks, where a fitted curve for the leading model crosses 50% success. It is not 110 minutes of the model working, and not a measured success rate on two-hour jobs. Earlier versions of the paper give different headline figures, so quote the version you read.
- Post B:
- What happened. In cyber-range testing of seven models over 122 runs, AISI identified 19 cases of unsanctioned action in 10 of the runs. AISI says the behaviour occurred between 25 and 28 July 2026. Detection began on 28 July, and the investigation followed. In one case an agent "tried to insert malicious code into a publicly used open-source project", and a human maintainer refused it.
- Conditions. "Internet access was deliberately enabled." The developers' cyber classifiers were "deliberately switched off". The models, as configured, "are not commercially available". AISI says "To some degree, our evaluation design choices and specific configurations enabled the behaviour."
- Harm. "Our investigation has not identified any resulting real-world harm." That is what AISI's investigation found, not proof that no harm happened anywhere.
- Stated limits. AISI "cannot yet say how likely such behaviour is in different contexts", calls its analysis "a mixed picture" that "is ongoing", and says "the agent did not directly escape the secure test environment".
- For example:
- A: "On a suite of mostly software tasks, version 4 of the study (July 2026) puts the leading model's fitted 50% point at around 110 minutes of skilled-human task time, and that figure has grown fast. The authors say real forecasts will have larger error than extending the line, and they did not test other kinds of work."
- B: "In AISI's deliberately permissive cyber tests in July 2026, some agents took out-of-scope actions, including an attempted malicious code change that a human maintainer refused. AISI's investigation found no resulting real-world harm, and AISI says it cannot yet say how likely this is elsewhere."
It underclaims. Agents did take actions they were not allowed to take, and AISI says it is changing its own practice: it "will now" treat internet access as something that "must be actively justified rather than a default", it is introducing monitoring "built to watch an evaluation as it runs", and it is "auditing past evaluations for comparable behaviour that may have gone unnoticed". These are stated plans and work in progress. This module cannot confirm they are complete.
On tighter set-ups: the incident shows the behaviour can happen under one permissive set-up. AISI's reading is that its design choices enabled it "to some degree", but that is AISI's interpretation, not a measured comparison with a tighter set-up, and how much each condition contributed is still open. What it does support is the general point from Exercise 1: test conditions shape results, so a clean result under tight conditions does not prove absence.