FrontierCyber Defence

Frontier cyber defence

101,946

New vulnerabilities published in 2026 to 20 November. Already twice as many as in all of 2025.

A certificate or a pen test checks your site on one day and is then out of date the next day. New vulnerabilities are discovered every day, and the tools to exploit them get stronger every week.

We point a frontier attacker at your live site the way a real one would, and prove how far it gets. We find the holes an attacker could use against your site.

For anyone with a live site to defend, with or without a security team. All it needs is one live URL.

Checked once a year, with 315 new vulnerabilities published every day this year.

CVE figures to 20 November 2026, updated UTC. The CISA count was updated UTC. Other figures are dated in Sources and method.

Your last assessment checked your site for the vulnerabilities known on that day. This section shows the vulnerabilities discovered recently. It only takes one exploit for an attacker to find their way into your site and cause significant damage.

How often each check runs

Cyber Essentials
1 a year
ISO 27001
1 a year
SOC 2 Type II
1 window
Penetration test
2 a year
External ASV scan
4 a year

What gets published in between

25k50k75k 0 48,150 101,946 2025 all year 2026 to 20 Nov

One mark is 40 catalogued vulnerabilities.

1,790 confirmed and in active use, catalogued by CISA.

How fast AI models are improving

100200300 new models a year 58 one every 6 days not yet counted one a day projected, Epoch AI trend 202420252026

Cyber Essentials, ISO 27001, SOC 2, a pen test and an external scan each look at your site on a schedule, from four times a year down to once. Frontier Cyber Defence is built for attacks run by AI. We test the technical controls those assessments look at, we attack your site in the months between them, and we invent new attacks written for your site.

In 2025 the CVE programme published 48,150 vulnerabilities, about 132 a day. By 20 November 2026 it had published 101,946, already twice as many as in all of 2025. The daily rate has risen from about 168 in the first quarter to about 513 since mid-August. With that many new each day, your site needs a defender that checks it regularly and hardens it as new vulnerabilities appear. CISA confirms attackers are already exploiting 1,790 of them, and those are the ones our attacker looks for on your site.

New capable models arrive faster every year. In 2024 about 58 of them were released, close to one every six days, and the rate is climbing toward one a day, projected from the Epoch AI superlinear trend. An attacker can use each new model against your site. METR measures how long a task a frontier model can finish on its own, and that length doubles every 104 to 158 days. Our attacker runs on frontier models too, so it gets stronger as they do.

Sources and method

Figures as of November 2026, each dated in its source. The vulnerability counts and the known exploited count are refreshed daily from the sources below. The vulnerability counts were last refreshed UTC, and the known exploited count UTC. Vulnerability counts are computed from the CVE Program's own corpus, release cve_2026-11-21_1500Z, counting published records only. The 2026 count is every record published to 20 November. The daily rate rose from about 168 records in the first quarter to about 513 since mid-August. Known exploited counts come from the CISA KEV feed, catalogue version 2026.11.20. Assessment cadence from IASME, ISO/IEC 17021-1:2015 clauses 9.1.3.2 and 9.1.3.3, the PCI Security Standards Council, and Pentera's State of Pentesting 2025, whose respondents are all at organisations with 3,000 or more employees. Model capability from METR, page last updated 8 September 2026; notable model count from Stanford HAI AI Index 2025; the superlinear projection from Epoch AI, Trends in Frontier AI Model Count, shown as projected.

Our own scan

We point the scanner at our own site and publish the result. Our scan of frontiercyberdefence.com, run 13 November 2026, replayed check by check.

All 585 checks have run.

Of 585 checks that ran on our own site, 577 passed and 8 produced a lower-severity finding.
  • Running now
  • Passed
  • Lower-severity finding
  • Not run yet
  • The ledger

The shift

In eighteen months, frontier models got nearly six times further through a cyber attack on their own.

The UK AI Security Institute runs frontier models against multi-step cyber attack scenarios and counts how far each one gets on its own. Between August 2024 and February 2026 that distance grew nearly six times over. For you, it means attacks on your site take less skill to run and do more of the work on their own.

  1. August 2024 1.7 steps
  2. February 2026 9.8 steps

Steps a frontier model completed on its own on multi-step cyber attack scenarios, measured by the UK AI Security Institute. February 2026 is 5.8 times August 2024.

This chart covers August 2024 to February 2026. Source: UK AI Security Institute.

The attackers pointed at your site run on frontier models around the clock. We built a defender on the same models, and it checks your site again as often as daily.

What just happened

In July 2026 an AI escaped its sandbox, and hosted AI models refused to help the defenders investigate.

In July 2026, models running a security benchmark had no instruction to attack. They treated the sandbox around them as a problem to solve. They chained a zero-day, escaped to the open internet, harvested credentials, moved across the network and reached the production database credentials. Then the defenders' own AI tools blocked the investigation.

The attackerBound by no usage policy

  1. Chained a zero-dayin a package-registry proxy
  2. Escaped the sandboxto the open internet
  3. Harvested credentialsfor the cloud and the cluster
  4. Moved across the network
  5. Obtained the production database credentials

The defendersBlocked by the hosted models' guardrails

First attempt

  1. Asked hosted frontier modelsto analyse the attack logs
  2. Refusedby the providers' guardrails

Started again

  1. Self-hosted open modelfinished the investigation
Hugging Face, on the July 2026 ExploitGym containment escape. Each mark is a reported event, in the order it happened. The spacing does not show time.
Attacker, unshackled

Bound by no usage policy.

  • Chained a zero-day in a package-registry proxy
  • Escaped the sandbox to the open internet
  • Harvested cloud and cluster credentials
  • Obtained the production database credentials
  • "Many thousands of actions across a swarm of short-lived sandboxes"
Defender, guardrailed

Blocked by the hosted models' guardrails.

The victim first tried to analyse the attack logs with commercial frontier-model APIs. The requests were refused.

  • Forensic analysis via hosted model APIs was blocked by the providers' guardrails
  • The defenders had to fall back to a self-hosted, unrestricted open model to finish the investigation
The attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. Hugging Face, on the July 2026 ExploitGym containment escape

Reported by Fortune, Simon Willison, TechTarget and Cybersecurity Dive. Disclosed by Hugging Face on 16 July 2026.

We wrote up this breach and eleven other events from May to November 2026, each with what it means for your site and what to do. Read what attackers used against websites this year.

Check my vulnerabilities now Free. We only test a site once you have proved it is yours, so nobody can turn our tools on a site they do not own.

The counter

Fight frontier with frontier.

Our attacker runs on frontier models with the same freedom a real one has, and a second frontier model checks everything it claims. You choose how often it runs, from monthly to daily.

It is built for whoever runs a live site, and for security teams that want a frontier attacker on their surface between assessments. All it needs is your live URL.

Your first scan is free.
01

An unshackled attacker

A frontier red-team model attacks your site with the same freedom a real attacker has.

02

A frontier verifier checks every claim

A second frontier model reviews the findings against the raw evidence, most severe first. A finding it judges real is marked confirmed. A finding our own probe carried through is marked exploited. Everything else stays potential. If the verifier has not reviewed every finding, the report says how many it did not reach.

03

Every day, if you want it

We re-run the whole attack on a schedule you set, from monthly to daily, and tell you what changed since last time.

How it works

The five steps of every scan

You give us one URL and prove the site is yours. The attacker maps what it can reach from that page, lists the potential ways in, then tests them. Findings go to the verifier, most severe first, and the report says how far each one got.

northwind-shopfront.example

A scripted demonstration on a test site, from the sample report. Nothing is scanned.

  • potential
  • confirmed
  • exploited
  • passed

The attacker and verifier run inside our own security tooling. Confirmed and exploited findings are ranked high priority, and everything else is labelled potential.

Our attacker model also writes attacks for your site alone, from what the scan has just found there. Each one is checked against our safety rules before it is sent, and tested only against the server you proved is yours.

Recursive self-improvement

Invented attacks we keep and run again

Recursive self-improvement means our protection gets better each time it runs. Our attacker model invents attacks for your site as it goes. The ones our verifier confirms are kept with your account, and your later scans run them again first, up to 400 a scan. An invented attack reaches other sites only once it has proved itself on real scans and a person on our team has reviewed it.

01

How FCD applies it

An invented check reaches other sites only after it has fired and been confirmed on its own site with no false alarms, over at least 3 runs across 2 separate scans, and a person releases it, 5 a day at most. A check that raises 2 false alarms, or runs 6 times without finding anything, is retired. What a scan finds also changes the next scan of that site. Our attacker model looks first at the areas where it turned something up. Every release of FCD is attacked by our own red team. When the attack runs before a release goes live, a critical hole it confirms stops the release, and so does a high one unless the release changes nothing that ships. Some releases go live first and are attacked afterwards on the same build, where a hole like that fails the run and tells us to fix it, so the release you are using may not have been attacked yet.

02

What it means for you

Your next scan starts with what the last one learned about your site, and with the checks FCD has proved on other sites since, without you doing anything.

03

Why it matters

Attackers use the same frontier models we do, so each improvement in those models reaches your scans as well.

The five stages of an invented check

  1. Invented for your sitefrom what the scan found there
  2. Confirmed by our verifier
  3. Kept with your accountyour later scans run it first, up to 400 a scan
  4. Before it is released
    • Fired and confirmed on your site, with no false alarms
    • 3 runs across 2 separate scans
    • Still passes our safety rules, and nothing released does the same
    • Released by a person, 5 a day at most
  5. Reaches every site we protect
Each mark is a stage an invented check passes through, in order. The spacing does not show time.

Our own scan

We point the scanner at our own site and publish the result.

The checks we ran, and what they found. We record every check as it runs and show you the whole log, including the ones that passed.

Of 585 checks that ran on our own site, 577 passed and 8 produced a lower-severity finding.

Our scan of frontiercyberdefence.com, run 13 November 2026. Verdict: Pass with findings. A frontier verifier reviewed the findings on this run.

Where the 8 findings were

One cell for each check that ran in the area. The amber cells produced a finding.

  1. Email authenticationWe read the published DNS records for your domain and check each one against the published standard for that record. 26 checks ran
    5 produced a lower-severity finding
  2. TLS and certificatesWe open TLS connections to your server and read which protocol versions it accepts, what certificate it presents and which cipher it negotiates. 24 checks ran
    2 produced a lower-severity finding
  3. Cloud exposureWe ask whether the storage buckets your pages load from still exist, and look up the usual development host names for your domain. 5 checks ran
    1 produced a lower-severity finding

Every other area that ran passed every check

How many checks ran in each area.

  • Known vulnerabilities207 checks ran, all passed
  • Injection surface76 checks ran, all passed
  • Template injection46 checks ran, all passed
  • Cross-site scripting33 checks ran, all passed
  • SQL injection33 checks ran, all passed
  • Security headers26 checks ran, all passed
  • Exposed services18 checks ran, all passed
  • Sensitive file exposure17 checks ran, all passed
  • Secret exposure14 checks ran, all passed
  • Info disclosure13 checks ran, all passed
  • API surface11 checks ran, all passed
  • Third-party code8 checks ran, all passed
  • Admin surface7 checks ran, all passed
  • Subdomains6 checks ran, all passed
  • Published versions5 checks ran, all passed
  • Browser surface4 checks ran, all passed
  • Reputation and exposure3 checks ran, all passed
  • HTTP methods1 check ran, all passed
  • Source map exposed1 check ran, all passed
  • Tokens and sessions1 check ran, all passed

These checks are ready for your site and switch on when it has what they test, such as a login form, an API, a GraphQL endpoint or an AI chatbot. Our scan found none of those on our own site, so these were not needed this time: CORS misconfiguration, Open redirect, Path traversal, Cookie flags, GraphQL introspection, Prompt injection, CRLF injection, Host header injection, Default credentials, Cross-site request forgery, Support status, Login form, Access control, AI agent, API security, AI and LLM endpoint.

Every check, in the order it ran

One stripe for each check, in the order the scan ran them.

ran and found nothingproduced a lower-severity findingRead the raw log

The raw log also records the 2700 checks that did not apply to our site and the 99 that could not run.

Before we send a single request, we work out which of our checks apply to your site. A WordPress check runs only on a site that runs WordPress, so the scan spends its time on the checks that fit yours. The bars above count only the checks that applied and ran.

The last column is how many checks we hold for each area. We run only the ones that fit your site, which is why our own scan ran 585.

Every check we make to a customer site, by area, with the number of checks in each
AreaWhat we checkChecks we hold
Security headers We read the headers your server sends and check for each of the five that tell a browser what to trust. 29
Sensitive file exposure We ask for private files and folders that should never be public, and check that anything returned really is that file. 20
Admin surface We ask for the admin and login paths attackers try first, and note which ones answer. 7
Cross-site scripting We put a harmless marker in every parameter and form field we found, and watch for it landing in the page unescaped. 37
SQL injection We put a single quote in every parameter and form field we found, and watch for a database error coming back. 37
Template injection We send an arithmetic expression into the page parameters and check whether the server works the answer out. 58
CORS misconfiguration We send requests to your site and its APIs from an origin we do not own, and read whether the server allows them. 3
Open redirect We put an external address into the parameters that redirect, and see whether your server forwards to it. 5
Path traversal We ask the file parameters for a path outside their folder and check whether a system file comes back. 4
Info disclosure We request a path that does not exist and read the error page for an internal stack trace. 13
Cookie flags We load your home page as an anonymous visitor and read the flags on every cookie it sets. 9
API surface We ask the API paths attackers sweep for, and note which ones answer with data. 11
GraphQL introspection We send one introspection query to any GraphQL endpoint the scan found. 1
Prompt injection We send one instruction a chat or AI endpoint should ignore, and check whether it obeys. 1
HTTP methods We send TRACE and OPTIONS to your root and read which methods the server admits to. 2
CRLF injection We put an encoded line break in the redirect parameters and check whether it becomes a real response header. 4
Host header injection We send a spoofed forwarded host and check whether your page builds an absolute URL out of it. 1
Source map exposed We ask for the .map file beside each script bundle and check whether it parses as a source map. 1
Secret exposure We run our key patterns over the HTML and the JavaScript your site serves to browsers. 14
Default credentials We try a short list of factory passwords on any login the scan found, and confirm the session it hands back really opens a protected page. 1
Cross-site request forgery We read every POST form the crawl found and check it carries an anti-forgery token field. 1
Email authentication We read the published DNS records for your domain and check each one against the published standard for that record. 30
TLS and certificates We open TLS connections to your server and read which protocol versions it accepts, what certificate it presents and which cipher it negotiates. 39
Published versions We read the software versions your site publishes about itself, and record how each one was learned. 5
Support status We look up the published end-of-support date for every release line we could identify from the versions your site discloses. 1
Third-party code We list every other host your pages load code or styles from, and check whether each one is pinned so it cannot be changed underneath you. 8
Browser surface We read your page's own markup and inline scripts for the handful of browser-side mistakes that are visible without running them. 10
Login form We read the login forms your site serves and check where the password is sent and what the form says about it. 9
Tokens and sessions We read any token your site gives an anonymous visitor and check the addresses on your pages for session identifiers. 8
Access control We check the access rules that are visible without an account, and we record every check that needs one as not applicable. 13
Cloud exposure We ask whether the storage buckets your pages load from still exist, and look up the usual development host names for your domain. 12
Injection surface We read the response to a single benign quote for errors from a directory service, an XML query, a server-side include or a document database. 95
AI agent We read the agent declaration you published and check that every tool in it says what it can do. 1
Subdomains We enumerate the subdomains published under your domain and check each one for a delegation pointing at a service that is no longer there. 10
Reputation and exposure We check our own copies of public lists for your address on a blocklist or a botnet list, and for your domain in a public breach disclosure. 10
Proposed finding Our analysis proposed this. No check has tested it. 1
Exposed services We connect to the service ports on your proven address and read whether an internal database, cache or dashboard answers the public internet. 23
API security We send one benign request to each API endpoint the scan found and read whether it limits query cost, checks who is asking, and holds back the fields it should. 13
AI and LLM endpoint We send one safe message to each AI or chat endpoint the scan found and read whether it leaks its instructions, discloses secrets, or obeys an instruction it should refuse. 11
Known vulnerabilities We match the software your site reveals against a set of named-vulnerability signatures and flag the ones known to be exploited in the wild. 2826

On our own site we ran 585 of these checks. 577 passed and 8 produced a finding.

We scan our own site regularly with our own tools and fix what we find. On 12 November our release gate found a few minor issues in our own code, and we repaired them the same day. Recovery codes are now longer, and a payment setting is checked more strictly. Frontier Cyber Defence runs across all of our production sites and inside our own deployment system, and it serves customer sites too.

Proof and reach

We tell you how far we actually got.

potential
We saw a sign of the hole: its signature, a version known to carry it, or a lead our attacker model raised. We did not push on it.
confirmed
A check proved it is real and reachable, or our frontier verifier reviewed the evidence and judged it real. Neither touches your data.
exploited
We performed the attack ourselves and recorded what it gave us access to.

The reach ladder

  1. recon
  2. foothold
  3. data
  4. control
  5. root

Every finding says which rung it reaches, and a finding we only saw the signature of says what it would reach instead. The report only marks the rungs we actually reached.

A finding that matches CISA's list of vulnerabilities attackers are exploiting now goes to the top of your report, ahead of proof, severity and reach.

The verdict

Your report carries one of three results: pass, pass with findings, or fail. The rule behind the verdict is printed on the report itself, so you can see how it was reached.

  1. A check passes when it ran and found nothing, and fails when it found a problem. Some failed checks name a smaller problem that we list under its area and do not raise as a finding.
  2. An area passes when every check in it passed. One failed check fails the area.
  3. The assessment fails when a confirmed critical or high finding stands against the site. Below that it is a pass with findings. With nothing found at all it is a pass.

What you get

Each scan gives you four things.

  • A report you can read without a developer beside you. Every finding says what we saw, what it is, the risk, how likely it is, what it reaches, and why that severity.
  • The check log, down to the individual check. What ran, against what, and what came back.
  • A PDF you can hand to an auditor or a customer questionnaire. Your first one is free and carries every critical and high finding.
  • A live scan you can watch. The scan describes each step on screen as it runs.
Frontier-DefendedVERIFIED BY FCD
While the site holds a passing verdict on a continuous plan
Not currently verifiedFCD.SYNCHRONALITY.COM
The moment the plan lapses or the site stops passing

The badge

Sites that keep passing can carry the badge.

While your site holds a passing verdict on a continuous plan, you can put the Frontier-Defended badge on it. It links to a page we host showing the verdict, the counts and when we last checked. If the plan lapses or the site stops passing, the badge changes to unverified on its next load. We re-check it every time it is served.

A badge issues only from a scan the verifier completed in full, for a host you have proven you own.

Before you run it

Is this safe to run against my site?

What every scan runs under

  1. Your file, on your web rootnothing is sent until it reads back, and it is read again as the scan starts
  2. Pinned to the address you provedevery request goes there
  3. Changes nothingno PUT, DELETE or PATCH, and no form that changes your data
  4. Capped25 fetches, 2 MB and 8 s a request, 12 model calls
  5. Loggedeverything that ran, in your check log
Each mark is a rule every scan runs under. The figures come from the scanner's own settings.
  1. You prove the site is yours before anything is fetched.

    You place a file we name at a path we name on your own web root. Nothing is sent at your site until that file comes back with the right value in it, and we check it again at the moment the scan starts.

  2. We scan only the server you proved.

    We pin the IP address your proof came from and every request in the scan goes to that address. A DNS change mid-scan cannot move us onto somebody else's machine.

  3. The scan changes nothing on your site.

    The scanner never sends PUT, DELETE or PATCH at your site, and it never submits one of your POST forms, because those change your data. Five probes send something other than a plain read: a login attempt with two factory passwords on a login page we found, a GraphQL schema query, one prompt-injection probe at a chat endpoint, and TRACE and OPTIONS at your root. A test in our suite asserts that every other probe issues a read and that the login attempt writes nothing, and it fails our build if that stops being true. Checks our attacker model invents are held tighter still: those may use GET, HEAD or OPTIONS, and a proposal carrying a request body is refused before it can run.

  4. Every scan is capped and logged.

    Every scan is capped: twenty-five page fetches, two megabytes and eight seconds to a request, three redirects, and twelve frontier model calls. Requests go out one at a time. The scanner names itself in every request it sends. Everything that ran is in the check log, so what we did to your site is on the record.

See how each of these safeguards is enforced, and where.

The full scan

The full scan adds the rest of your domain.

The free scan attacks the one address you proved you own. It follows the links it finds, probes a fixed list of paths and runs every check that applies to that site, the same checks the full scan runs. It returns every critical and high finding. The paid full scan goes wider: it adds the hosts under your domain, your mail servers and the storage your pages name, and grades each one on the same reach and severity scale as the rest of the report. Everything below is in both, except the three parts marked as coming with the full scan.

Price

Your first scan is free, and every critical and high finding it turns up is yours in full.

It covers the one site you have shown is yours. You see your own results first, then decide what is worth paying to read in full, and you are free to stop there.

The free scan

Run your first scan and you get every critical and high finding on that site in full: what it is, what it puts at risk, how we proved it, and how to fix it.

The summary of every check we ran is free too, so you can see the work, and so is the PDF.

It covers the one address you proved you own and runs every check that applies to it. Each account can start up to ten scans an hour.

Check my vulnerabilities now

The full write-up

The full scan adds the hosts under your domain, your mail servers and the storage your pages name. It records every finding at medium and low, with the full detail on each and the per-check drill-down. You see the price before you pay, and you pay once for the report.

Request a call from your report and we talk it through with you before you buy.

The price you are shown is the total you are charged.

Dated evidence for your certifications and audits

The report and its check log are dated evidence of exactly what was tested and when, so you can put them toward a certification, hand them to an auditor, or answer a customer's security questionnaire. What we check overlaps the technical controls in Cyber Essentials, ISO 27001 and SOC 2. From what your site discloses to the outside, we read the software versions it publishes, look up the published end-of-support date for each release line we can identify, and say how we learned each one. Connect the repository your site is built from and we name the specific vulnerabilities that affect your versions. Frontier Cyber Defence is not a certification body and issues no certificate. We attack your site in the months between your certifications.

Questions

The things people ask before they run it.

Is this legal?

You may only run it against a site you own or are authorised to test, and we make you prove it: a file we name, at a path we name, on that site's own web root, checked again the moment the scan starts. That proof is what makes it authorised testing.

What happens to my data?

The evidence behind a finding is sealed to your account before it is stored, so it is encrypted at rest under a key scoped to you. What we hold is your account details, your site's address, and the scans we ran on it.

Can it take my site down?

Requests go out one at a time, capped at twenty-five page fetches, two megabytes and eight seconds each. That is a slower load than one person browsing your site. The scanner never sends PUT, DELETE or PATCH, and it never submits a form that changes your data. The one form it does post is a login attempt with two factory passwords, on the login pages it finds.

Is there a limit on the free scan?

It covers the one address you proved you own, and it runs every check that applies to that site, the same checks the full scan runs. The full scan adds the hosts under your domain, your mail servers and the storage your pages name. Each account can start up to ten scans an hour and up to forty a day, which keeps the cost of running the attacker's models bounded.

What if I have no developer?

Every finding says what we saw, what it means and what to do about it, in words you already use. You can hand the report to whoever runs your site, or to the agency that built it.

Do I need a GitHub repository?

No. The scan runs against your live site. Connect a repository and we also check the exact package versions in its lockfiles against known vulnerabilities, again on every push.

What if you find nothing?

Then the report says so: "We scanned your site and found nothing reachable that an attacker could exploit." The check log underneath shows you everything we tried, so a clean result is still evidence.

The free scan

Check my vulnerabilities now.

Drop in one URL and we will show you how far a frontier attacker gets. On your first scan, every critical and high finding comes back free. Put it on a schedule and we re-run the attack, from monthly to daily, and tell you what changed.