US cyber incidents are flat since Claude Fable 5, an Epoch and Ipsos poll finds, but it cannot see AI

Have the new cyber-capable AI models made ordinary Americans less safe online? A poll published on October 7, 2026 by Epoch AI says the answer, so far, is that nobody can detect a change. The finding is real. So is a limit the authors state themselves: the poll "does not identify incidents caused by AI."

Epoch's analysts, Caroline Falkman Olsson and Yafah Edelman, asked the same question of a representative sample of US adults twice, three months apart. Their headline is that US adults are no more likely to face cyber incidents than when Claude Fable 5 launched.

The data

Both waves ran on the Ipsos KnowledgePanel Omnibus, a probability-based online panel recruited through address-based sampling, with estimates weighted to US adults aged 18 and over. The June wave ran June 12 to 14, 2026, with 1,021 respondents. The September wave ran September 18 to 20, with 1,024.

The results, as Epoch reports them:

• Any cyber incident in the past 12 months: 46% in June (90% confidence interval 44 to 49%) and 45% in September (42 to 47%). The table gives 46.2% and 44.6%.
• Frequent AI users, daily or weekly: a similar flat reading, 53% in June and 51% in September, on samples of 233 and 215.
• Spam or convincing messages: 32.2% in June, 32.1% in September.
• Personal data leaked in a hack: 19.8%, then 18.2%.
• Account break-in: 7.9%, then 6.0%.
• Service stopped by an attack: 7.8%, then 6.1%.
• Malware: 5.7%, then 5.0%.
• Ransom lockout: 2.1%, then 0.9%.

Epoch found no incident type that rose in a way it could reliably detect. Several fell slightly, but the intervals are wide and the authors do not claim an improvement.

Why June is the baseline

The June poll ran days after Anthropic publicly released Claude Fable 5 and gave vetted partners limited access to Claude Mythos 5, per Epoch. So this is not a before-and-after comparison of the release. It compares two moments in the early months after it, and it cannot say what the first few weeks looked like. Epoch also cites a spike in serious vulnerability disclosures in June and a further rise in July, which it treats as the reason to ask the question at all.

That context helps explain the interest. In the weeks since, the cyber side of the race has moved. NIST's CAISI said on September 17 that Z.ai's GLM-5.3 is "the most cyber-capable open-weight model released to date" and lags the US frontier by about four months. Anthropic's Frontier Red Team wrote on September 29 that GLM-5.3 produced end-to-end exploits in 50 of 410 attempts on ExploitBench, against 56 of 410 for Claude Mythos Preview. We covered both in our piece on GLM-5.3's cyber capabilities. If capability is spreading, a household poll is a slow instrument for noticing.

What the poll can and cannot show

Epoch lists its limits clearly:

• No AI attribution. The poll asks about incidents, not their cause. A phishing message written by a model looks the same to a respondent as one written by a person.
• Independent samples. Each wave polls a random sample "rather than tracking the same individuals," and only about 40 respondents took part in both. It measures population-level change, not change in the same people.
• Power. "Smaller changes may go undetected due to lack of statistical power."
• Overlapping recall windows. Both waves ask about the past 12 months, so the two periods overlap heavily.
• Self-report. Recall error, misclassification and non-response bias that weighting does not correct are all possible.

Add one limit of ours. The poll covers households. The incidents that worry defenders most, such as attacks on businesses, hospitals and infrastructure, would barely register in a survey of 1,000 adults. The October 6 report we covered, on a Mythos-found flaw exploited within a day of its write-up, is the kind of event this instrument cannot see.

What the next wave needs to show

A third wave would add the most if it used a larger sample and a panel that follows the same people, so that the roughly 40 repeat respondents become several hundred. It would also help to ask what the respondent believes caused the incident, even though self-reports of cause are noisy. Epoch has not said whether it will run further waves, and we have not seen a schedule. Until then, the June to September comparison is the only public household-level reading we have on whether the Fable 5 generation of models changed daily online risk.

What this means for the debate

Both sides will use the number. Those who think cyber risk from frontier AI is overstated will cite the flat line. Those who think the danger is ahead will point out that the models were new, access was gated and the poll measures the wrong population. Both would be partly right. The poll is a good baseline, which is most of its value: if incidents do climb after more open-weight cyber-capable models are released, Epoch now has two earlier points to compare with.

Our take

Treat this as the first honest public tripwire, and a quiet one. Nothing detectable happened to ordinary people in the months after Fable 5, which is worth knowing and is not a safety result. We would watch the next wave, any larger sample, and enterprise incident data that can attribute causes. The useful question is no longer whether the models can do harm, which NIST and Anthropic's own researchers have documented, but whether anyone is measuring it well enough to notice when they do.