Time-to-exploit just went negative. Now what?

Zero-days are now found at machine speed, and patching can't keep up. Here's how confidential computing buys back time: hardware-encrypted enclaves and attestation keep a compromised host from ever reaching your data.
https://www.anjuna.io/blog/time-to-exploit-just-went-negative-now-what

Buying back time with confidential computing in the Zero Days era

TL;DR Zero-days are now found at machine speed. Patches are not, so the strategy of patching faster not only breaks production but also continues to leave you exposed. Here’s where confidential computing comes in. Workloads run in hardware-encrypted enclaves, and attestation releases secrets only to verified, unmodified code. A compromised host, hypervisor, or cloud operator gets the machine but not the data.

Can you really patch faster than machine speed?

In January 2021, Qualys disclosed a heap overflow in sudo (CVE-2021-3156) that had been in the code since 2011. One bug, nearly ten years in the wild, found by a human research team. We wrote about it at the time, and the argument we made then has aged well: patching is necessary, but it cannot be the entire strategy. It has to sit alongside controls that protect you before the patch exists, and before anyone knows the bug is there.

That was the old pace. Recent headlines tell a different story. The pace is quickening… rapidly. Since April, Anthropic’s Claude Mythos Preview has reported thousands of zero-days across every major operating system and browser. Firefox 150 shipped with fixes for 271 vulnerabilities found in a single evaluation pass. Claude Fable 5.1 made source-level vulnerability discovery a generally available product in September. Moonshot’s Kimi K3 does the same as a free, open-weight download. And on November 5, 2026, OpenAI released GPT-6 Cyber, which found 81 new high-severity flaws in Firefox and built working exploits for 948 Linux kernel bugs.

The fixing side has not kept up. By late May, Anthropic had disclosed 1,596 vulnerabilities to open-source maintainers; 97 had been fixed. OpenAI’s own Patch the Planet dashboard shows fewer than a quarter of its findings closed.

The window after disclosure is gone too. Time-to-exploit has gone negative: the Zero Day Clock now records exploitations dated prior to CVE publication, 37 of the latest 97 cataloged, with another 60 on publication day. When the clock reads below zero, you cannot know when your exposure began.

So every security team is hearing the same instruction: patch faster. But patching at machine speed is how you break production, and it still leaves you exposed to the bug nobody has found yet. There is a better posture, one where moving carefully is safe. The rest of this post is about how to get there.

If root can read it, root can steal it.

Why does a bug in sudo or a kernel carry so much weight? Because it lands on an assumption the whole stack is built on: whoever gets root gets the data. Security has retired assumptions like this before. For twenty years we trusted anything inside the network perimeter, until enough breaches started on the inside and zero trust replaced it. Root access to data is the same kind of assumption, and it is due the same treatment. Admin access has always meant the ability to query the database, dump a process’s memory, lift the keys out of it and walk away. Encryption at rest and in transit does not change that. The moment data is in use it is plaintext in memory, and memory is readable by anyone with sufficient privilege on the machine: your own admins, the cloud provider’s operators, and anyone who reaches the hypervisor through a bug nobody has patched yet.

Secret Stores move that problem rather than remove it. Something still has to authenticate to the vault, and that first credential, the one the industry calls Secret Zero, ends up in an environment variable, a config file or a container image, in the clear. The chain always bottoms out somewhere, and wherever it bottoms out, root can read it.

We have made this argument before, in The Privileged User Power Differential and Who Guards the Guardians?, so the short version will do: if root can read it, root can steal it. A zero-day that gets an attacker to admin is a zero-day that gets them the data. That link is what every emergency patch cycle is racing to protect, and it is the link worth breaking.

Shrink the blast radius, not expose the whole stack

Confidential computing severs the connection between privilege and data. Applications run inside a trusted execution environment, a hardware-enforced enclave where memory is encrypted at the CPU level. Only that trusted workload can read the data in unencrypted form, and that holds even against the most privileged software and users on the infrastructure, your own privileged admins, the cloud provider’s operators, an insider whose access was always broader than the job required, and any unauthorized user who reaches that level of privilege through a vulnerability. That includes Secret Zero. The one credential that unlocks all the others is no longer sitting on the host waiting to be read.

That changes what a root compromise is worth. The attacker still has the host. They can list processes, consume resources, take the machine down. What they cannot do is read the memory of the workload, lift the keys out of it, or alter the code and have it keep running. That is confidentiality and integrity together: the data stays unreadable, and the code stays the code you deployed. Confidentiality comes from the hardware encrypting memory at the CPU level. Integrity comes from attestation: Anjuna Policy Manager releases secrets only after cryptographically verifying that the application is running inside a genuine enclave and is unmodified. Tampered code does not get the keys, so it has nothing to steal. It also closes the Secret Zero problem. Rather than planting a bootstrap credential and trusting that nobody finds it, the workload’s attested identity is the credential: secrets are released only to code that has proven in hardware that it is the right code, in a genuine enclave. There is no Secret Zero to issue, rotate or steal.

Anjuna Seaglass is what makes this practical on the infrastructure you already run: AWS Nitro Enclaves, AMD SEV-SNP, Intel TDX, and NVIDIA GPU workloads, across AWS, Azure, Google Cloud, and on-premises. It packages your existing container into an Anjuna confidential container and runs it unchanged. No rewrite, no enclave SDK, no splitting the application to work around vsock. The local socket is an enclave’s only channel in or out. The Anjuna Seaglass confidential runtime delivers first-class security with minimal performance overhead.

Let’s be clear about the boundary. A trusted execution environment protects the workload from everything beneath it; it does not protect the workload from itself. An SQL injection flaw in your own application is just as exploitable inside an enclave as outside. What changes is the size of the problem: instead of defending the whole stack against every kernel, hypervisor and container-escape zero-day, you are defending your own code and its interfaces. It also cuts lateral movement short. A compromised host next door is no longer a path into the enclave’s data, and each confidential workload is its own isolated trust boundary rather than one more machine on a network where root is transitive. That is a much smaller surface, and a far more tractable one to patch carefully.

What confidential computing buys you: Time (with a capital T)

On October 20, 2026, the Wikimedia Foundation took Wikimedia Cloud Services and other systems offline as a precaution. Two memory-safety flaws in PHP, CVE-2026-109618 and CVE-2026-109619, allowed unauthenticated remote code execution in every version from 5.4 to 8.5. They had been in the code for more than a decade. They were found in August by a researcher at California using OpenAI’s GPT-5.6 Sol, and independently by Wiz. Details leaked the day before the coordinated release, so the fix shipped early and everyone else moved at once. Finland pulled its government web services. The Dutch cyber security agency took down its own site. Agencies told organizations that could not patch in time to consider going offline instead.

That is the real cost of a zero-day for most teams, and it is rarely the breach. It is the choice between staying up while exposed and going down to be safe, made at speed, with incomplete information.

As mentioned above, a flaw in PHP is a flaw in the application runtime, and code execution inside an enclave is still code execution inside an enclave. Wikimedia’s own explanation of the shutdown names the shared resource model as the problem: a vulnerability in one hosted project threatens the others. When each workload runs in its own enclave, a compromise is contained to that workload, and the host it sits on is not a path to anyone else’s data.

For the much larger category of zero-days that live below your application, the change is bigger. Last week supplied one. OpenSSH 10.6p2 shipped on November 5 with a fix for SSHole (CVE-2026-123158), an authentication bypass that lets anyone who can reach the SSH port log into certain accounts by supplying a password between 257 and 1024 characters. Proof-of-concept code appeared about four hours later. GreyNoise recorded the first successful exploitation in the wild on November 7, roughly forty hours after the patch. CISA added it to the Known Exploited Vulnerabilities catalog on November 9 with a remediation deadline of November 12. The configuration is specific: it affects OpenSSH built against the musl C library without PAM, which is Alpine Linux’s default. Anyone on glibc, using PAM, or running key-only SSH was never exposed.

The post-exploitation behavior is the part that matters for this argument. Reports describe reconnaissance, calls to cloud metadata services, searches for stored credentials and SSH private keys, and changes to authorized key files. That is an attacker who has the host doing the only thing left to do: hunting for Secret Zero. This is where confidential computing changes the outcome. The workload’s memory is unreadable from the host, and there is no Secret Zero on disk to find. The attacker gets the machine and nothing that is on it.

It is not a one-off. In July, Broadcom shipped VMSA-2026-0006: CVE-2026-47876, an out-of-bounds write in ESXi’s VMXNET3 adapter rated 9.3, alongside two critical vCenter flaws at 9.8. The ESXi bug is a VM escape. An attacker with admin rights on one guest executes code on the hypervisor, and from there reaches every other VM on that host. There was no workaround. Patching meant a host restart, so Broadcom classified it an emergency change and told customers to vMotion workloads off and reboot clusters in a rolling wave.

While the OpenSSH and VMware bugs are wildly different, they both end in the same place: an attacker with privilege on infrastructure outside the enclave. That is the whole category confidential computing neutralizes, and it is much larger than two CVEs — including container escapes, host kernel flaws, a compromised management plane, a cloud operator. A shell on the host and an escape to the hypervisor both stop being routes to the data, because the data was never readable from either. The patch still matters and the restart still has to happen, but it can happen in next week’s maintenance window rather than tonight. You patch on a schedule you set, after you have understood the vulnerability and tested the fix, instead of at 2 a.m. against a countdown.

None of this is permission to run vulnerable software. An unpatched system is still an unpatched system, and the backlog still has to clear. What confidential computing buys is the thing teams never have during an incident: time to think and act.

Slow is smooth, smooth is fast

This hamster wheel does not slow down. Better models meet more code every quarter, and the discovery curve is pointing one way. The patching curve is not going to catch up, because the constraint there is people, testing and change windows, and none of those scale the way inference does.

So the honest question is not how to patch faster. It is: what your posture looks like when you cannot - which is often the case. Confidential computing is not the whole answer to that. You still need good security hygiene like zero trust, least privilege, monitoring, and a patch program that works. Confidential computing is the layer underneath all of them: the one that decides whether a compromise anywhere in your infrastructure becomes a compromise of your data. Get that layer right and the rest of your controls are defending something, instead of being the only thing standing between an attacker and your plaintext crown jewels. The next zero-day is already out there; the question is only what it costs you.

That is where the time (with a capital T) comes in. When you know your data security holds, you can read the advisory properly, test the patch, schedule the restart and ship it clean. Slow is smooth, and smooth is fast.

Anjuna Seaglass runs your existing containers in hardware-backed enclaves across AWS, Azure, Google Cloud and on-premises, with no code changes. We offer a free 30-day trial. Contact us to find out more.

More like this
Get Started Free with Anjuna Seaglass

Try free for 30 days on AWS, Azure or Google Cloud, and experience the power of intrinsic cloud security.

Start Free