Talk to us

The hundred-year flood is another day at the office

Christian Have
LINKEDIN

Christian Have

Christian leads product, engineering, architecture, quality, and AI and data at Helm. He has more than 20 years of experience in security across law enforcement, healthcare and global enterprise, including work with the Danish National Police, and was previously CTO and Chief Product Officer at LogPoint. He holds a degree in information security from the IT University of Copenhagen, has completed executive programmes at MIT, and advises and speaks at international security forums including Europol, Interpol and NATO.



What the government CISOs told me

In the summer, a couple of weeks after Anthropic announced its Mythos model, I spent a day with a group of government CISOs. Patching critical vulnerabilities within a month was already hard for them, and they were now being asked to do it in hours. The incidents they had actually handled had come through vulnerabilities that were years old, patched elsewhere long ago, in systems that were not on anyone's list of internet-facing assets.

My own experience is the same: zero-days are rarely behind the big incidents. The causes are authentication bypasses, file uploads and credential leaks that are two to six years old, unpatched, and one day exposed to the internet or connected to it by a contractor. GreyNoise found the same in its 2024 data: 40 percent of the vulnerabilities exploited that year were from 2020 or earlier, some from the 1990s.

Volume and velocity

Since those CISOs started their careers, attacks have grown in volume and velocity, and software now does the targeting and the exploit writing. On 31 August an operator pointed a few hundred AI agents at PaperCut, a print-management server used by schools and offices, with two bugs that had been disclosed four days earlier. The agents wrote the exploits, pulled a target list from an internet scanning service and broke into their first real victim in under four hours. By the time GreyNoise had counted, 395 organisations in 48 countries were compromised, close to half of them schools, and at twelve of them the agents had domain admin, reached through noPac, a Windows bug Microsoft patched in 2021. The schools were on the list because they were reachable and ran PaperCut.

Mandiant's mean time to exploit was 32 days after disclosure in 2021 and seven days before disclosure in 2025. Its median time between an attacker's first access and the hand-off to a second group was eight hours in 2022 and 22 seconds in 2025. To me that second number means that access to a compromised company is now auctioned off by machine before anyone on the defending side has seen an alert; even the ransomware economy is algorithmically traded. The fastest ransomware families encrypt about 25,000 files a minute, by Splunk's measurement, and nine in ten ransomware deployments run outside business hours, in Sophos's incident data. A large organisation writes an incident like that off as a bad quarter; for SMEs, the World Economic Forum cites estimates that 60 percent of those that fall victim to a cyberattack close within six months.

The limits of patching faster

For the part of an estate that updates itself, browsers, phones and software a vendor runs for you, faster patching is already happening and is the right answer. For the rest it is the wrong place to put the effort. The emergency-patch decision was always expert work. In a top-tier vulnerability programme, every Patch Tuesday fix is analysed within days. The patch is diffed against the old binary, the bug is rated for how easy it is to exploit, intelligence sources are checked for use in the wild, and the likelihood of weaponisation is forecast at a day, three days, a week and a month. The decision at the end is whether to push an emergency patch outside the agreed change cycle. In a large estate that costs millions a time in downtime, testing and staff time, so the call is made a handful of times a decade, and even with that effort it is sometimes wrong. Most organisations work from a severity score and a service level, and both assume an attacker who needs weeks to build an exploit. In June Anthropic's model wrote a working Firefox exploit within an hour of Mozilla shipping the patch. When Anthropic pointed the same model at open-source code in the spring, maintainers asked Anthropic to slow down its disclosures because they could not write patches as fast as the model found bugs, and by 22 May, 75 of the 530 high- or critical-severity bugs it had reported were fixed. In this year's Verizon DBIR the median time to fully fix a known-exploited vulnerability rose to 43 days, from 32 the year before. I think of this as threat exposure debt: known bugs in software you run, with no fix installed, accumulating faster than they are retired.

A faster cadence has its own cost; one faulty CrowdStrike update crashed about 8.5 million Windows machines in July 2024. And a large share of any public-sector estate cannot be patched in hours at any cadence: industrial controllers, medical devices that need re-certification, appliances nobody has on a list. The refrigeration controllers manipulated in hundreds of French supermarkets in October, and in Sweden, Belgium, Luxembourg and Estonia on 2 November, were end-of-life units that CERT-FR had told operators to take off the internet three weeks earlier.

How we readjust

Engineers size flood defences for the hundred-year flood, the one with a one percent chance of arriving in any given year. As the climate changes, that flood arrives far more often, and the people who run physical infrastructure have had to rethink what they build for. AI is doing the same to cyber: the incident that used to come once in a career now comes every quarter, and an organisation has to be run as if the next one is another day at the office.

For anything on the internet, treat a known bug as already exploited and limit what an attacker can reach from it: take management interfaces off the internet, segment so that one compromised host reaches as little as possible, and put a web application firewall or CDN rule in front that blocks the pattern while the fix ships. Cloudflare and the large hosting platforms had rules in place within a day of the PHP2Shell disclosure in October for the millions of sites behind them, and one of the PaperCut agents was stopped by a web application firewall. For anything that cannot update at machine speed, the goal is to make the vulnerability unreachable rather than to fix it, with a safe manual fallback.

What happens at the first malicious action has to be decided in writing before the incident, because no one on the defending side can act within 22 seconds. When files on a share start being encrypted from an account, the encryption is evidence enough. The risk owners can decide now that the account is cut off at once and the share isolated, with a person reviewing the incident afterwards. That decision belongs to the business, and it is easier to take in a quiet month than during the incident.

Recovery has to be measured. In Gartner's 2026 assessment of 307 organisations, Respond had the widest gap between rated importance and reported maturity of the six NIST functions, 3.9 against 3.0 on a five-point scale. Containment time and recovery time are the two figures a CISO can set, rehearse and report before any incident.

How Helm approaches containment

Helm detects encryption as it starts on file shares, storage and cloud platforms, isolates the account responsible and records what happened for recovery and reporting. The organisation sets the rules for that first response in advance, so they can be reviewed, rehearsed and audited.

See how it works: Contain. Control. Continue

Related reading