cns me / blog Cloudy with a Chance of Freefall
← Index | | 23 min read

The part we can see

AI Generated: My shelves are pretty full, but it’s the empty one that everyone is drawn to.

Bottom line up front

Earlier today Isabella Ward @ WIRED published a story about a vulnerability I reported to the Planning Inspectorate : a leaked credential granting access to their entire Microsoft 365 tenancy. It’s one of over 118 vulnerability reports I’ve submitted via the government’s Vulnerability Reporting Service since June, (funded by ControlPlane). So far, only 15 have been closed as resolved. 54 are currently ‘in progress’ (median 108 days after reporting), 14 of which are closures I still dispute. All the vulnerabilities are the sort of ‘bread and butter’ ones listed in the OWASP Top Ten. The lesson to learn from this is not that we should close code, but that open code makes these types of vulnerability easy to identify and validate (most probably would be accessible via the live services anyway). Closing code is an example of security through obscurity and only serves to remove people who report vulnerabilities for free. The lesson, I think, is a financial one. HM Treasury rules incentivise building software as an asset rather than operating it, and most security fixes manifest in the operation of software. So, cheaper identification of vulnerabilities means longer queues. Keep coding in the open, invest in operating and make testing a habit.

I’ve listed 5 findings here, as they are published on ControlPlane’s website. These include the Planning Inspectorate credential, 3 Department for Education services and a Scottish Government open data portal. Everything else remains abstract, as it would be irresponsible to publish a concrete description of a vulnerability that has not been addressed. I’ll be adding more as they are fixed and published. ControlPlane are a security consultancy that provides assurance and AI security services, so they have a vested interest in some of the things I’m saying, the ControlPlane CEO is also the OpenUK CISO. I’ve published these numbers so that others can draw their own conclusions from them. Security researchers: please point out where I’m wrong.

The flaw WIRED covered

The Planning Inspectorate had a Microsoft Entra ID app secret (full-blown client ID + tenant ID) lying around in a public GitHub repo for 18 months. It had been force-pushed off the branch, but GitHub still served up the old commit if anyone asked. Removing a secret from a branch is like removing a biscuit from a photograph of the table. We can all still remember the biscuit and some of us even made notes about it. The credentials granted full read/write access to everything the Planning Inspectorate had in Microsoft 365, including the Preservation Hold document libraries used to store evidence on legal hold. I told WIRED it was ‘god-tier’ access. That’s what the credential allowed, not what I did with it. I’ve listed the access it had, but I didn’t download any file contents and I’ve demonstrated write access by probing without making any changes. I reported this at 12:10 UTC on the 26th August 2026 and gave it a CVSS of 10.0 (maximum).

And then, a model response. The Planning Inspectorate’s monitoring had identified my activity before I reported it. About 20 minutes after reporting the issue, the secret was no longer valid as it had been revoked. The Planning Inspectorate has since told WIRED it verified that no data breach occurred. There was a technical briefing that afternoon, and the team went on to review all the other repositories they had and restored them as public. This is how it is done: monitor, react, learn and remain transparent. Although the report is still open on HackerOne, it doesn’t count towards the 15 resolved. But the issue was closed within the hour.

Kitchen table with laptop, egg timer and crumbs, a cable trailing out of the window
AI Generated: The biscuit is gone but we all know where it was.

Built in the open

Putting that code back in public view is important because it’s a celebration of over 10 years of building in the open. The UK public sector has created around 27,000 repositories in total at some point. The current cross-government leaderboard shows 24,280 public repositories from 212 organisations, but there are thousands more that have since been deleted or closed. The combined total of all the government organisations on GitHub from every other country is 33,868. That’s a different way of counting, but together those numbers suggest the UK has around 42 per cent of all the public sector repositories on GitHub. I wrote the scraper for that leaderboard back in October 2018. I’ve written before about why coding in the open is a good thing, and also about Asking 24,500 Repositories a Question where I asked the estate questions in English.

Coding in the open is a policy, not an option. The Service Standard point 12 explicitly states that new source code must be open, while the Technology Code of Practice instructs teams to share their code and use open-source. The GOV.UK Design System is the manifestation of that policy. Launched on 22 June 2018, by March 2022 (according to a report from GDS) it had been reused over 3,300 times, its GOV.UK Frontend codebase had been downloaded more than 2.7 million times, and pages using it loaded approximately twice as fast with half the amount of code. Since then it has continued to evolve, with a brand refresh starting in June 2025. If you fix something in a component that is used across multiple services, then it gets fixed in every service.

Blueprint cross-section of a building with transparent upper floors and a queue of boxes at the base
AI Generated: The upper floors are all glass and the lower floors are all waiting.

The local government version is called LocalGov Drupal . This started at Brighton & Hove City Council in 2018 and has since spread to Croydon Council and beyond, thanks to the Local Digital Fund. It’s operated by councils for councils via an Open Digital Cooperative board. More than 50 councils use it, according to their website, and they claim you can save up to 80 per cent on the cost of a new website. Councils who could never afford to build their own content management system can now share one they can trust.

There’s also CyberChef. A GCHQ analyst started this in 2014 during innovation time, and it was open sourced in November 2016. It bills itself as a Swiss Army knife of cyber stuff: a web app for encoding, decoding, encrypting and data analysis. The 2021 case study from GCHQ cited over 75,000 monthly users and 1.4 million unique users to date. An intelligence agency gave away the implements of the trade for fundamental security analysis work, and the world ran with it. I love it when a spy agency releases a tool that lets you drag and drop steps into a recipe and out pops the data from the other end like a very earnest sausage machine.

I was only able to do this research at this scale because of open code. I’ve been part funded by ControlPlane since June to review public sector code using a combination of Claude Code (Opus and Sonnet, mainly Sonnet). Testing with a much cheaper Gemma class model locally identified many of the same issues. The reviews were of public code, and didn’t skip any guardrails despite being on the Cyber Verification Programmes to have access to Mythos and Daybreak. Any real-time checks ended once a finding was confirmed, and any changes were reverted. In total, it reviewed 349 codebases across 33 organisations, identifying 1,094 issues in at least 223 of them, at an approximate cost of $50 worth of AI tokens per service. That’s the cost of getting some candidates, not verified reports, which took much longer to do. The adage about open-source that it has many eyes was basically fulfilled by one paid pair of eyes.

Why the basics keep breaking

Software engineering is hard, and software enigneering best-practice is harder than it looks. But the types of issues that have been identified (broken access control, hardcoded credentials, injection, lack of authentication) are all well-known and have been for a long time (they are all on the OWASP Top Ten). Government security frameworks demand regular testing, and government departments put it in their contracts (one Department for Education one demands an independent IT Health Check pre-launch and then at least yearly). This is a habit that goes back to the Public Services Network, which has seen its plug pulled since 2020, but the need for assurances didn’t go with it. A typical health check costs thousands of pounds per service. How many of these services have had one, I don’t know. I’m not anti-health check, just that there is a place alongside them for a cheap, always-on scan.

Manuscript illustration of a long queue of scribes carrying scrolls to a sealed door
AI Generated: Everyone has something to say, but nobody has the key to the money.

But why does it seem that the UK public sector finds so much of this so hard? The January 2025 State of Digital Government review provides the best answer. It found that the funding model was skewed towards new policy programmes and discrete project funding, with 28% of central government technology estates being legacy and more than a quarter of the most-at-risk legacy systems not having any remediation funding. The 2020 and 2021 spending reviews had only 43% to 46% of new digital spend being funded from day-to-day budgets, compared with a benchmark of 78%. However, those legacy numbers are from before the August 2026 change where the definition of legacy has shifted to be about burden and risk, and includes systems that cannot be effectively secured, regardless of age. That change is likely to increase the numbers of what is considered legacy.

“Many have highlighted that it is easier to get CDEL (capital funds) than RDEL” (State of Digital Government review, DSIT, January 2025)

The answer is actually set out in the Treasury’s own rules. Capital budgets are for new investment. Resource budgets are for day-to-day spending, and the HM Treasury budgeting guidance makes clear that control over day-to-day spending is of paramount importance. Because custom software is an intangible fixed asset, creating it will (normally) be a capital expense, but the costs of ownership, including maintenance, will be a resource cost. Government departments can choose to reallocate money from resource to capital budgets (so long as that money is not ring-fenced). However, they can never reallocate from capital to resource. Finally, there are the fiscal rules. The Charter for Budget Responsibility sets an aim to borrow only for investment and a requirement to have a current budget surplus by 2029-30, with balance or surplus thereafter. Capital is also subject to a cap, but it is treated more generously. A sustained increase in day-to-day spending would require either higher taxes, cuts in other areas or a guaranteed share of existing budgets. None of those is politically easy. ‘And also, maintenance’ is a phrase that has never been greeted with a standing ovation at a party conference. The hall remains seated and someone checks their phone.

But software gets bought, then put out to pasture and left to rot (or patched if you’re lucky), before eventually getting replaced. The project team often leaves at go-live. It’s a bit like employing a removal company to move your sofa into your new home and then they drive away, leaving you to find the stopcock for yourself by trial and error. A rebuild that makes the asset better can be capitalised (depending on the accounting policy of the department), but day-to-day fixes cannot. In a previous post I had built a map of the public sector’s software bills of materials. Revisiting this map in June, I found that 402 out of 755 public-facing repositories had at least one Critical or High severity dependency vulnerability. That’s 16,568 vulnerabilities in total, with a handful on CISA’s known exploited list. Most of those are build tools that don’t get used during a live request, but nobody is doing any pruning.

Open under sink cupboard with a stopcock, a lone sofa in a bare room beyond
AI Generated: You don’t think about the bucket until the day of the pipe.

It’s not just about funding though, it’s also about thinking in projects rather than products (and software as a service is already part of day-to-day budgets, so the capital bias is mostly around bespoke development like much of the open estate). The Performance Review of Digital Spend, published in March 2025, identified that there is often a lack of funding for maintaining and improving services, suggesting pilots that mix change and run funding. Following the 2025 Spending Review, Government Digital Service and HM Treasury also ring-fenced important cyber and legacy programmes. None of this (as far as I can tell) currently funds the intake of vulnerability reports though.

To me, this is the point where the economics of cheap discovery run up against the budget rules. AI has made it about $50 of tokens (at list price) per service to find flaws, but pretty much every step between finding and fixing is on the resource side of the ledger: triaging, patching, rotating a secret. So, cheaper discovery does not reduce the length of the queue, it expands a queue that can only be addressed by day-to-day money. (See an earlier post about replacement dates for why I think whoever moves a date must be able to pay for the implications on the service that remains running.) The budget rules are one of the reasons they often can’t. The numbers tell you where the queue is.

Whiteboard map with sticky notes queuing at a narrow gate and scribbled question marks
AI Generated: There’s a lot of post-it notes lining up for that one tiny doorway.

The numbers

I submitted 118 reports from 18 June 2026 to 4 October 2026, across 20 different HackerOne programmes, with 100 of those coming in June and July. Of those 118, 76 had a CVSS of 7.0 or higher, with 15 of those being 9.0 or higher. There are also another 8 that I had marked as Critical in the title, but have no CVSS recorded, and 2 that were never rated. I set 105 myself, with 13 being set by the programme teams (again, none above High). The 118 are a subset of the 1,094 issues that the review identified. I manually reproduced and confirmed all of the 118 before submitting, so they are skewed towards the more serious end. They are a sample, not a census.

Currently, 15 of the reports are marked as resolved (12.7%). 54 are in progress, of which 34 are not yet triaged, 6 have been triaged and 14 have been closed without fixing but are still under dispute. Another 49 have been closed without fixing in a way that I accept (26 as informative, 15 not applicable and 8 as duplicates). However, some of those may well be correct. Reports are only considered to be in dispute if I have chased them in the last month, so my chasing activity creates movement between the two categories. If you preefr to consider all closures as final, then 40 are still open and 63 were closed without fixing.

Of those that were fixed, the ones that were fixed quickly were very quick. The 15 resolved reports had a median of 13.7 days, with 8 fixed in 14 days or less. However, 6 were more than 80 days old and the average was 49.3, so that median is only representative of the ones that were fixed. Here are three published examples to illustrate the range. The Scottish Government’s open data portal contained a long-lived administrator token in the public JavaScript. With a CVSS of 9.8, Critical, it was fixed within 2 days when the owner of the service decommissioned a service that was no longer required. The Department for Education ’s Register of Training Providers used a good authorisation library but missed the check for several administrative actions, meaning a standard staff account could deactivate administrators. With a CVSS of 8.3, High, it was fixed within 14 days. The Get help buying for schools service allowed anyone who knew the identifier of a request to amend it, including the contact email. With a CVSS of 6.5, Medium, it was also fixed within 14 days.

The problem is with everything that doesn’t get to an engineer. The aim of the service is to triage reports within 10 working days, but as of the 6th November, 54 reports were still in progress, with a median age of 108 days (48 were more than 90 days old). This varies quite a lot by programme: one that got 32 of the 118 reports hasn’t managed to fix any, and has 29 of the 34 not-triaged reports in their queue. 15 of the 21 not-applicables had no programme activity after the initial day, which suggests an automated close after about a month (but no-one has confirmed that at programme level). Some bugs got fixed but people didn’t tell anyone. One was 11 days after I reported it, and 11 days later the report got closed as informative because I hadn’t shown impact. I can’t always prove that my reports led to those fixes.

A tall in-tray of unopened buff envelopes on a worn wooden desk beside a rubber stamp
AI Generated: The stamp is eager but has not yet been invited to work.

The 5th published case is a more subtle example of this sort of issue. In the Department for Education's service aimed at teachers applying for qualified teacher status, code that connected a GOV.UK One Login sign-on to an existing application was falling back to searching by email when it didn’t find a matching identifier, then overwriting any existing identifier. I only tested with one mailbox and an account that belonged to the research team. I reported this on 26 June 2026, with a fix visible in source by the 4th August. The issue was marked as resolved on the 23rd September, 50 days after the fix was visible.

Now, to be fair to the triagers, I made their job harder by submitting 100 reports in 2 months. That’s a lot of reports for anyone to handle. The use of AI generated junk is clogging up disclosure programs (curl ended their bug bounty in January 2026, for example, but theirs was a flood based on paying money which this service doesn’t do). I can only speak for myself, but I manually verified each report before submitting them. However, that’s not something a triager can know upfront. I don’t know if the central queue is underfunded, under prioritised or understaffed. Publishing the age and triage times would help us know that. If it’s a funding issue, it should be treated as a standing cost.

I also did some work on looking for secrets. The GH Archive has nearly all the public pushes of code to GitHub since 2011, but they’ve been having a hard time keeping up this year (in the hours I checked in September they only had about 22% of the push events), so my numbers are probably an underestimate. With that, I’ve triaged 7,616 potential secrets from commits that had been orphaned by force pushes in 813 UK public sector repos. Of those, 13 were active credentials (most verified with a single read-only check) compared to 7 across the rest of the governments I checked (that’s 13/20 from a country with ~42% of the repos). However, that’s a tiny sample, I was mostly looking in the UK, and the places that are most measured look the sickest (what epidemiologists call detection bias).

The part we cannot see

Some people are locking down their code. In response to the emergence of AI models like Anthropic’s Mythos (7th April 2026), NHS England released guidance to make all their code repos private by default from the 11th May (I can’t find any inidcation they have since changed their mind). Terence Eden wrote about this, pointed out that it didn’t make anything more secure, and helped ensure all their repos were backed up. 3 days after the deadline, GDS and DSIT released their own guidance to set against it in favour of transparency. I’m still on that side of the fence. If you want to know how many people are making their repos private, the UK Gov Mirror (by Jonty Wareing ) shows the number of deletions rather than repos being made private. There are 3,234 deleted repos from 90 organisations, with 2,054 of those being this year. Of the 1,278 deletions in October, 903 were from a single organisation, for reasons as yet unexplained.

“Openness should remain the default posture, with closure used sparingly and deliberately.” (GDS and DSIT, AI, open code and vulnerability risk in the public sector, May 2026)
Terrain map half hidden by fog, with a compass rose whose needles point several ways
AI Generated: It’s a lot easier to navigate when you can actually see the fog edge.

If it’s public, there is a copy somewhere. Audris Mockus’ research project World of Code is attempting to create an index of all public version control. When I checked 738 government organisations against it, I discovered 7,686 repositories that are indexed but no longer exist on GitHub. Of those, 3,686 are from UK organisations. Scanning the 3,641,335 versions of files in the deleted UK repositories, I found 94,778 credentials-shaped strings.

Any respectable threat actor is monitoring our public outputs. When Comparitech put AWS keys in public GitHub repositories, they were abused within a minute or so. Deleting a repository stops people who report vulnerabilities as a public service from seeing them easily, not people who cloned it ages ago. So it’s not really open and buggy vs closed and bug-free, it’s open and checked vs closed and unchecked. I can’t measure the closed estate, but I’ve no reason to believe it’s any less buggy, just that people can’t view from the outside and sharing in gov being what it is, neither can most the people on the inside..

But why close the code? There is certainly an argument for keeping certain things private, and the guidance does acknowledge that there can be valid use-cases for making limited exceptions. However, making new code private by default is an attempt to reduce future exposure, which is fair enough. However, closure is only a delaying tactic if attackers need the source code, and I’m not convinced they do for most of these vulnerabilities. Blanket closure is security through obscurity. This will not fix the vulnerabilities that currenntly exist or the costs to fix them, and the guidance itself says that making code private should not be a replacement for other security controls. It’s the ostrich effect. Karlsson, Loewenstein and Seppi have shown that investors check their portfolios more frequently in rising markets than falling ones, so making a repository private is the equivalent of an organisation not opening its statement.

Split image of a Victorian clerk locking a safe beside an unwatched modern server rack
AI Generated: The old clerk would not understand the server, but he would see the open door.

The moment after access

There’s a moment when a finding appears and my heart misses a beat. For a second the possibilities come faster than the rational thoughts: could I view something private? could I impersonate someone? Then comes the conscious decision to resist, a decision that I have to make with every finding. I can’t rely on my previous record of good behaviour. I wrote 3 versions of a post about that moment (I called it The moment after access) but never published any of them. This is the one I’m going to.

Article content
WarGames: "Do you want to play a game"

In the movies, people don’t just stop. In the 1983 film WarGames, a teenager alters his grades on his school district computer system, then hacks into a military computer via a backdoor password that the creator had left behind and, well, nearly starts a nuclear war. Both WarGames and 1995’s Hackers are films that I have fond memories of from my youth, and I never thought I’d still be quoting them in later life, but a lot of the issues I’ve identified are not that dissimilar to that backdoor; credentials and access methods left lying around. Hackers just made it all seem sexy, what with the rollerblading and a young Angelina Jolie in what was her first big studio movie, before she went on to become one of the most recognisable sex symbols on the planet.

Defence is not sexy. All an attacker needs to do is be right once, but defence (team blue) needs to be right all the time to keep services secure. This is what Google’s security leaders term the ‘defender’s dilemma’. They claimed in February 2024 that AI could solve this, but my data suggests that AI can only solve it if people pay to act on the information it provides. The costs of defence are high and the work is, largely, not very exciting: applying patches, rotating secrets, checking logs, responding to alerts. Good monitoring can help to level the playing field a bit (see the Planning Inspectorate’s early detection of my activities), but you don’t make a film about rotating a secret. It is that day-to-day stuff that the budget constraints make hardest to fund.

The hard thing for a researcher is what happens after you stop. One service recently closed down one of my reports because I hadn’t shown live impact. But showing impact on the live service would have required me doing something I was not allowed to do. So I asked for a permission to test a small number of things and didn’t proceed without it. I hate having to carry that responsibility and doubt with me. The service’s policy does not grant permission to break the law, it just promises to make it clear you were compliant if anyone does accuse you of it. The government’s promised protection for good-faith research under the Computer Misuse Act has not yet made its way on to the statute books. Information bias, as Baron, Beattie and Hershey put it, is valuing a test that the outcome of cannot influence a decision. If you need proof as a service running something, either give the researcher permission or do the test yourself.

Then there’s the chasing. I sent 3 rounds of chasing in late October: 5 on the 21st, 15 on the 27th and 30 on the 28th (all to different reports). The service limits researchers to chasing the same report once every 14 days, and many of these would have been months since the previous chase. One was a simple yes/no question on a single report, and this was the 4th time I’d asked in 2 months. As of the 22nd July, even after multiple reports had been resolved, HackerOne’s own getting-started checklist still told me to submit my first valid HackerOne report. Each unanswered chase is a lesson to a researcher that their next report is not worth the time and effort.

Now, that’s been a fairly thankless job to be honest (my best guess is 200-400 hours, closer to the upper end: up to 10 weeks full time, with almost no day going by where I didn’t do some). The scanning was the easy bit, the hard work is in the triage, chasing, evidence and write-ups. These services are built with public money, bugs and health checks included, but the Vulnerability Reporting Service doesn’t incentivise reporting them, and it says pretty clearly on GOV.UK there is no bug bounty. (Some countries do pay: when Singapore launched its bug bounty programme in 2021, it was offering between US$250 and US$5,000.) I’m listed on the acknowledgement pages for the UK and Berlin, so I’m grateful for that. But this is neither exciting AI, nor a Mythos-style discovery, nor a brand new product. Most people just don’t care about broken access control on a service they’ve never heard of.

A notebook, laptop and mug of tea on a desk by a rainy window at dusk
AI Generated: There’s always a still evening like this one after a discovery, with the light on.

What I would like to happen

None of this would have happened if it weren’t for ControlPlane, who kindly paid for the work from the first report through to this one. Nor would it have happened without Andrew Martin , founder of ControlPlane . I’ve known Andy for a long time and we remain firm friends through a shared love of tech, security, red wine and London steak houses. Thanks also to the engineers who got things fixed in minutes and hours, the team at the Planning Inspectorate for showing how it should be, Terence Eden for defending open code, and Jonty Wareing for keeping score. Hopefully this data can help you all make the case.

Now, here is what I’d like to see happen. Keep the code in the open, it is the number one priority and the only part that can be properly scrutinised. Pay to run by including an allowance for triage and maintenance each year as part of the cost-base of the service, ringfencing that money against in-year raids. Measure the time from report to triage with as much rigour as time to fix. Make it a thing to do continuous testing. Secure by Design already encourages teams to continuously monitor the effectiveness of their controls through manual and automated testing in addition to NCSC’s Web Check. I’d like to see National Cyber Security Centre / Government Cyber Coordination Centre or another body accredit AI testing services in the same way that CHECK and CREST accredit penetration testers. Purchasing one should become an accepted part of doing business, and a regulated industry can grow up around it, which will help ControlPlane (and its competitors). I’d want that anyway.

Woodcut poster of a marble hand beside a blindfolded Greek temple with server racks in ruins
AI Generated: You can’t hide a temple by closing its doors.

Next, once the intake is funded, set the date. In a world where fridges are turning into ovens (supermarket fridges across Europe are apparently overheating due to ‘attacks’), ongoing testing should become a requirement for organisations who provide critical services and services that are funded by the taxpayer. That should start with central government departments and their various arm’s length bodies. AI has made it many times cheaper to find flaws. If the funding and people to fix those issues don’t also appear, every new (cheaper) scanner just adds to the queue. Closing the code will continue to look like the solution, when it is really just a blindfold. The part we can actually see is not the problem. It’s the only part of the problem that people are looking at.

(Views in this article are my own.)

🦩