Mark Johnston (@cognitivecyber) header
Mark Johnston (@cognitivecyber) profile picture

Director - Cyber Security Leader @ Google Cloud. Tweets are my own, we need to collaborate to defend. Safety in our digital economy is key. #TechnoCISO

Singapore
Joined January 2011
Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)Photo by Mark Johnston (@cognitivecyber)
Imagine a world where most of the code we write and maintain is free of vulnerabilities. I’ve long been skeptical whether formal methods were practical, but with this new research it’s clear that LLMs will make this possible. With a basket of solutions — proofs, pre- and post- submi agentic code scanning, and autonomous redteaming — we will build a more secure future. Congrats to Pushmeet and the team on the research. (And yes I manually typed those emdashes!)
Announcing a major milestone achieved by my team @GoogleDeepMind: we have proven key security properties of the machine code underlying the Signal messaging protocol. This is the first time the actual production code of a widely deployed protocol has been formally verified – by any method, with or without AI – and at this depth. More than a research breakthrough, it’s a prime example of how frontier AI can be a force multiplier to counter the speed and scale of modern cyber threats. Formal verification offers a higher standard of security; however, it’s extremely complex and time consuming, which is why in many cases it’s considered to be impractical. Similar efforts could take teams decades of combined effort. We did it in a month’s time. Using AlphaProof Code, Google DeepMind's new agentic framework for formal verification that runs on LLMs including Gemini Carbon, we verified the Signal protocol down to the actual code running on devices.  This milestone lays the groundwork for a future in which vital systems, from medical device firmware to power grid software, ship with mathematical proof that their protections work. Reaching that future will take continued research from our community. In that spirit, we're open-sourcing our formal specifications and verification artifacts so others can build on this work. On a personal note, this moment is of deep significance to me. I started my own research journey more than two decades ago as an intern working on checking the consistency of complex software systems. Seeing this grand challenge made practical with modern AI is extraordinary. I'm extremely proud of my team members: Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger,  Karn Seth, Jonathan Katz, and Charles Sutton for taking on this grand challenge and delivering, and grateful to the Signal team for their partnership. deepmind.google/blog/formal-…
9
11
1
67
5,737
Spoiler alert… this is not the defense mantra at Google, and we do have some ideas about machine-speed defense. Check out our first paper on Beyond Zero. Early experiments and prototypes are showing a lot of promise this could be a good direction to go, further supported by some of the standards work. blog.google/security/going-b…
Calif CEO @XorNinja says cybersecurity has no playbook for stopping 1,000 AI agents attacking a network simultaneously, even if defenders can detect them: "Attackers always adapt faster and earlier than defenders. Even the AI companies themselves don't have great security." "Imagine you enter a cage fight, but you are blindfolded and the other guy stay in the dark, and every once in a while they would come out and punch you and then disappear." "How could you win that fight?" "I was at Google and that's basically the mantra at Google. In security you have three things: prevention, detection, and response. And Google accept that prevention gonna fail at scale." "You can detect them, but does it help you if you find that there's maybe a thousand agents hacking you right now inside your network? How can you contain them without a full shutdown of your system?" "This actually requires defense to come up with a new paradigm. I haven't seen anybody come up with a good idea here. We're not there yet." @calif_io
9
22
1
161
18,055
'One of my favourite cyber security maxims is Halvar Flake’s observation that  “All offensive problems are technical problems, and all defensive problems are political problems”' ncsc.gov.uk/blogs/one-does-n… (spotted at @unpromptedconf and reshared here)
17
110
7
506
24,451
Mark Johnston retweeted
CVE-2026-98432 is a good example of something that really concerns me about the future of cybersecurity. ubuntu.com/security/CVE-2026… The vulnerability was introduced in October 2015, more than 11 years ago, by a performance optimization in the Linux TCP/IPv4 networking stack. One of the most fundamental components of an operating system used by millions of servers. Now Anthropic reports that Claude discovered this flaw and developed a working exploit that allows remote code execution as root, without authentication, against a server with an MPTCP listener. anthropic.com/research/disco… Think about what that means. For decades, we’ve been building critical infrastructure out of enormous amounts of complex, memory-unsafe code. Humans wrote it. Other humans reviewed it, tested it and searched for vulnerabilities. All working with the same fundamental limitations: limited time, limited attention, incomplete visibility and an inability to reason about every possible interaction and race condition. And perhaps that’s what protected us to some extent. Attackers had the same limitations as the people developing and defending these systems. A vulnerability could remain undiscovered for 11 years, even in such a widely used component. Now AI is changing that balance. The ability to inspect code, identify subtle weaknesses and develop working exploits is becoming less dependent on a small number of highly skilled researchers. These activities can increasingly be scaled with compute, automated testing and tokens. And that’s what worries me about the next few years. We’ve accumulated decades of complex, memory-unsafe code in operating systems, network services, applications and libraries. Much of it has never been assessed with anything close to the capabilities that are now becoming available. We may have massively overestimated the security of our software simply because the people looking for vulnerabilities were as limited as the people who introduced them. That balance is breaking down, and I don’t think we’re prepared for what follows.
46
80
10
361
42,271
Mark Johnston retweeted
Some stats on #php2shell. In the 10 hours before it was publicly announced, across the @Cloudflare network we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post announcement, hundreds of thousands within 2 hours, and millions now.
285
3,256
675
22,557
2,284,419
Mark Johnston retweeted
Some magical progress towards "magic state" in quantum! Our @GoogleQuantumAI team has demonstrated "magic state cultivation" on physical hardware for the first time. The result is a 40x error reduction on our Willow processor vs. previous methods, achieving a state-of-the-art 99.99% logical fidelity. It's another step towards useful, error-corrected quantum computers to help solve the hardest problems. The result is published in Science today, watch the video below to see how it works
419
1,102
232
9,260
697,725
Friendly vulnpocalypse reminder: patch / upgrade asap. The Deep Clean is upon us.
"Several vulnerabilities" in the Linux kernel lwn.net/Articles/1097401/
2
3
14
2,463
Yeah this is a real take, right share at right time thanks for sharing @joedaroo
Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there.
1
1
3,659
Mark Johnston retweeted
Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there.
443
417
334
2,955
1,408,004
Mark Johnston retweeted
Can our TPUs survive and operate in space? Well, we're going to find out. Project Suncatcher is hitching a ride aboard @SpaceX's Transporter-18 mission, testing a prototype satellite built in partnership with @planet One small step for TPUs....
608
1,358
448
12,848
4,224,405
Mark Johnston retweeted
For 20+ years @ShaneLegg and I've discussed AGI’s potential impact on the economy, science & society. With the DeepMind Institute, we're expanding interdisciplinary research on key questions for the AI era. We hope it spurs the discussions needed to get the next steps right: deepmind.google/institute
340
627
127
3,972
628,506
Important work.
JUST RELEASED: Our finalized interagency report with @NIST shares technical recommendations for safeguarding the identity tokens & assertions that underpin single sign-on, federation, & application programming interface-based access scenarios. More 👉 go.dhs.gov/5uK
5
26
3,531
Mark Johnston retweeted
It’s done. My book, Scaling Security: The 7 Pillars of Elite Cybersecurity Programs, will be launched on February 23rd, 2027 and you can pre-order it now. amazon.com/Scaling-Security-… _______ We are all dealing with more complexity, more dependencies, more surface area, more risk, rapid technology change, and increasing attacker capabilities. At the same time, our businesses and societies have become ever more digitally dependent. What separates elite security programs from merely competent ones, is not simply better tools or bigger budgets. It is seven pillars that transform security leadership from reactive expertise into organizational leverage: setting leading indicators of performance, modernizing for inherent defensibility, prioritizing the high-impact 20 percent, amplifying people through structure and AI, architecting for resilience, running security like a business, and weaponizing speed as strategy. This book provides a leadership operating system for CISOs and other security leaders who need to move their programs from bespoke to scalable, to create organization-wide systems that hold under pressure - to move from the artisanal to the industrial.
3
7
1
40
2,094
Pssst reminder. Defense is going to win.
Congratulations to the CrowdStrike team on fine-tuning Nemotron 3 Nano 30B-A3B into a reasoning-enabled detection-triage classifier. What they did: • Input: Windows endpoint context, including process lineage, command lines, prevalence, ATT&CK mappings, severity, sensor metadata, and sometimes analyst context • Ground truth: analyst-assigned TP/FP labels • Training: prompt optimization, self-training, and RLVR using those labels as ground truth What does this mean for cyber defense? It demonstrates a locally deployable, specialized triage classifier. However, this is not a general-purpose cyber defense LLM, investigator, or threat-hunting agent whose primary goal is threat detection and retrieval. The model reasons over the context already provided. It does not retrieve new evidence, query security tools, correlate across data sources, or conduct an investigation. One result stands out: CrowdStrike reports that general-purpose models clustered around 55% to 71% accuracy on this task, roughly in line with an untrained NVIDIA Nemotron 3 Nano 30B-A3B. This looks different from our Cyber Defense Benchmark, where larger frontier models dominate the top ranks and Nemotron 3 Super places 24th out of 25 models. The results raise an interesting question: Would models excelling at threat hunting outscore smaller ones if they were given access to the environment logs? Cyber Defense Benchmark: simbian.ai/research/cyber-de… CrowdStrike research: crowdstrike.com/en-us/blog/t…
2
8
64
9,443
Mark Johnston retweeted
In an industry first, we’re piloting double-blind evaluations for frontier AI. By creating a secure environment where neither test prompts nor model weights are revealed, we can ensure external safety and performance evaluations of our models remain private, robust, and trustworthy. → goo.gle/3St2xan
122
105
48
857
369,307
So, yes, you noticed I've been blogging a lot lately. I am clearing my 2BLOG folder finally. Expect 2-3 or more fun blogs to land at medium.com/anton-on-security in the next 10 days....
2
2
26
2,312
In 2014 Google introduced BeyondCorp to build zero trust networks and systems. I'm super excited to be sharing our next major paradigm shift for the AI era: Beyond Zero. Learn how we are building a dynamic security framework that operates at machine speed: blog.google/security/going-b…
1
8
24
2,070
Mark Johnston retweeted
Or put another way Google is piling money into a business growing at 80% a quarter with a nearly $500B backlog of contracted workload to meet.
1
2
1
14
2,415
Mark Johnston retweeted
U.S. data centers require: 30x less water than U.S. golf courses. 100x less than California almonds. 200x less than U.S. home lawns. 2,500x less than U.S. crop irrigation. Who is protesting almonds? Golf courses? Lawns? Nobody. China does not feel threatened by your lawn.
403
1,582
124
6,966
218,492