Research
Device code phishing in Entra ID: what the sign-in logs actually show
Device code flow exists so a device without a browser — a TV, a CLI, a conference-room display — can sign a user in. The device shows a short code, the user types it on a phone, and the device receives tokens. Nothing about that requires the two to be near each other, which is exactly what makes it useful to an attacker.
What the sign-in logs record
Two rows matter, and they are not adjacent. The first is the authentication itself, logged against the client application the attacker chose. The second is the token redemption. In a legitimate flow both come from the same network within a couple of minutes. In a phished one the redemption arrives from the attacker.
SigninLogs
| where AuthenticationProtocol == "deviceCode"
| extend Redeemed = todatetime(CreatedDateTime)
| project UserPrincipalName, AppDisplayName, IPAddress, Redeemed, ResultType
| summarize Addresses = make_set(IPAddress), Attempts = count()
by UserPrincipalName, bin(Redeemed, 1h)
| where array_length(Addresses) > 1The query is deliberately loose. Tightening it on AppDisplayName is tempting and wrong — the application is the attacker’s choice, not a constant, and pinning it is how a hunt stops finding the next campaign.
Why MFA did not help
Conditional Access policies are commonly scoped to interactive sign-ins. The device code flow is not one. An account with MFA enforced everywhere a user would notice can still be taken over here, and the sign-in log will show the policy as not applied rather than as failed — which reads, at a glance, like nothing happened.
The policy did not fail. It was never evaluated. That distinction is the whole finding.
What to do about it
Block the device code flow for everyone who does not demonstrably need it, then allow it back by exception. In most estates the genuine need is a handful of shared devices and a few engineers, which is a small enough list to maintain by hand.