If you are responsible for security, last week’s warning was addressed to you. More than 150 of the largest technology, security and financial companies, OpenAI, Google, Microsoft, AWS and Anthropic among them, signed a letter that begins: “We have a limited window to strengthen cyber defenses.” AI-enabled attacks, they warn, will become far more widespread and sophisticated in the coming months; the status quo will not be enough. We agree. And one item on their list is ours to fix.
The weakness under the weaknesses
The letter names what has left systems exposed and asks for collaboration and higher standards. Most of that list, patching, permissions, misconfigurations, is hard work the industry knows how to do. The hardest item is the one none of it addresses: weak authentication. The industry has worked hard here too, from passwords to MFA and from MFA to passkeys, and attackers still get in. Proofpoint reported in December 2024 that almost half of the accounts taken over by bad actors had MFA configured.
Here is the shortcoming as we see it. The industry uses cryptography everywhere, the best technology digital security has, but it uses it in pieces, not end to end. Authentication produces a verdict (yes, this is the user); it does not protect the data exchange that follows. What carries the user afterwards is a token, which is only supposed to be difficult to copy. All the data is encrypted by the standard TLS protocol, yet the switch from the verdict to the data is a gap, and it is the weak link attackers, AI or not, exploit.
What AI changes
AI makes an attacker better at two things: producing deception a careful person will believe and finding the flaw in complicated machinery. It moves faster than defenses designed for human-speed attacks, and it probes everything where an attacker once had to choose. It can more successfully lure users into setting up the attack that is hardest to defend against: Adversary-in-the-Middle (AitM), where the attacker relays between the user and the real service, and both ends see a login that succeeded.
What AI is not good at is breaking cryptography. A properly designed, cryptographically atomic operation, where the authentication produces the encryption key that protects the data, addresses the issue: the AitM attack fails even if the lure succeeds.
The fix: a key instead of a verdict
Our login asks the user to notice nothing, and it is not assembled from separate pieces. It rests on a key inside one endpoint’s hardware that never leaves it. The Live Key represents the user, the device and local security policy all being met, and it disappears when any of those conditions stops holding. A password is worth something to anyone, anywhere; a stolen token works on any machine that presents it; this key is worth nothing to anyone who is not that user, on that endpoint.
The mechanism is not new, and it is not ours. It is mutual TLS, where both ends prove themselves, not only the server, and it has secured machine-to-machine authentication for decades. It never reached the ordinary login because it asked people to carry cards and handle certificates. What was missing was a client key that could stand for a person. The endpoint provides it.
Service providers, we need you for the session
We can close the login; we cannot on our own close what happens after it: the session between the user and the service. What we need is for applications to speak mTLS with the endpoint and the Live Key. Then login and session collapse into one mTLS session, with no user action and no login prompt left for AI to attack. Parts of this are already public: the LIT project on GitHub carries a reference implementation of the Live Key and mTLS authentication on Windows. We are looking for service providers and partners who are ready to build with us. We will provide further source code under agreement.
We knew it was coming: we published this before AI made it urgent
We first described the flaw in 2024, in “WinMagic Discovered a Flaw in TLS and FIDO,” and returned to it in “Did Your Login Pass or Fail?” Since then it has gone where it belongs: a formal proposal at the W3C in March 2026, drafts at the IETF, comments on the record with MITRE, and coverage from Authority Magazine to Mobile ID World. The flaw was there whether or not anyone exploited it. AI is exploiting it now.
Our part is not a roadmap: MagicEndpoint is deployed today, and the reference implementation is public. What remains is the work that cannot be done alone: independent testing of the login, and applications that speak mTLS with the endpoint.
If you want to examine what we have built, or build the session side with us, write to research@winmagic.com.
See our press release.
