Cyber PredictionsDaniel Miessler's predictions on AI and cybersecurity, 2016 to 2026

Better and better open models with no controls on themTons of new harnesses designed to help people do harm across cyber, organized crime, etc.

Capable open models inside skilled attack harnesses make AI a real security problem.

Every call, 2016 to now

When he said itFirst outside evidenceNo evidence yetHas happened Underway Not yetLinear time: equal years, equal length

2016

1 call

  1. No. 01 Consequences

    Attackers will increasingly use machine learning to qualify victims, triage fuzzing results and mine their scans for new targets.

    So the way this applies to infosec is like this: When attackers or defenders need to confirm a known thing, they might use Supervised Learning. And where they want to search for new ways to find attackers (or new victims) they will use Unsupervised Learning. … Expect both attackers and defenders to be using both with increasing frequency in the coming years.

    So the way this applies to infosec is like this: When attackers or defenders need to confirm a known thing, they might use Supervised Learning. And …

    Full quote

    Blog How Cyber Attack and Defense Can Leverage Supervised and Unsupervised Learning 

2018

1 call

  1. No. 02 Harnesses

    Small elite attack teams will build automation systems that crawl the whole internet and feed algorithms that pick targets, replacing thousands of human hunters.

    So elite attack teams will basically rig up extraordinary automation systems that constantly crawl and parse the entire internet, with special focus on certain targets, and then take all the data they find and put it into a format that can be consumed by various types of algorithms. … So what we'll have is the battle of the algorithms.

    So elite attack teams will basically rig up extraordinary automation systems that constantly crawl and parse the entire internet, with special focus …

    Full quote

    Blog The Future of Attack and Defense is AI 

2023

4 calls

  1. No. 03 Consequences

    AI-powered attacks will include faster phishing and automated exploit writing, reverse engineering and BEC, and attackers will out-use defenders with these models.

    Here's a fun question: who is going to be better at using ChatGPT and future models for cybersecurity attack and defense? Will it be the attackers or the defenders? My money is on the attackers for no reason other than them having higher numbers, more time, and more scrappiness. … Examples of AI-powered attacks will (and already are in some cases) include faster and better phishing campaigns, automated exploit code writing, automated reverse engineering, automated BEC and other social engineering, information warfare campaigns, etc.

    Here's a fun question: who is going to be better at using ChatGPT and future models for cybersecurity attack and defense? Will it be the attackers or …

    Full quote

    Newsletter Unsupervised Learning NO. 364 

  2. No. 04 Consequences

    AI will help attackers first: they will automate attacks for three to five years while defenders still lack the internal context to use AI well.

    AI will most help the attacker side first, and then it will help defenders more in the long-term. … In the early days of AI, attackers will be able to use AI to automate attacks while defenders still lack context about their environment.

    AI will most help the attacker side first, and then it will help defenders more in the long-term. … In the early days of AI, attackers will be able to …

    Full quote

    Blog Who Will AI Help More—Attackers or Defenders? 

  3. No. 06 Harnesses

    Dedicated hacker AI assistants will wrap attack and C2 frameworks, enumerate targets automatically and go all the way to launching live exploits.

    Performs automated target discovery and service enumeration on all discovered targets … GLitCH has access to 31 of the best commercial attack and C2 frameworks. … can go all the way to launching live exploit code at targets (with permission, of course)

    Performs automated target discovery and service enumeration on all discovered targets … GLitCH has access to 31 of the best commercial attack and C2 …

    Full quote

    Blog AI's Predictable Path: 7 Components 

2024

6 calls

  1. No. 07 Consequences

    AI's main edge in offense and defense will be moderate expertise applied to billions more targets and data points, not superhuman skill.

    To me this NSA story highlights that in both offensive and defensive security use cases, the main advantage of AI will not be its exceptional (superhuman) capabilities, but rather the ability to apply pretty-good-intern or moderate-SME level expertise to billions more analysis points than before.

    To me this NSA story highlights that in both offensive and defensive security use cases, the main advantage of AI will not be its exceptional …

    Full quote

    Newsletter Unsupervised Learning NO. 415 

  2. No. 08 Harnesses

    AI agents will sort through millions of targets, choosing what to attack, when and how, and attackers will adopt this far faster than defenders.

    One of the most powerful uses of AI—and especially agents—will be going through millions of targets and figuring out what to attack, when, and using which technique. … Red will move much faster than blue on this

    One of the most powerful uses of AI—and especially agents—will be going through millions of targets and figuring out what to attack, when, and using …

    Full quote

    Newsletter UL NO. 417: NSA’s Broker Buys, AI-Assisted Attacks, Companies Only Want Killers 

  3. No. 09 Harnesses

    Teams of agents loaded with captured tester methodology will find about 90% of typical web bug-bounty findings within two to five years.

    The way forward on automated hacking is this: 1) teams of agents, 2) extremely detailed capture of human tester thought processes, lots of real-world examples, and time. I suspect that in 2-5 years, agent-based web hacking will be able to get 90% of the bugs we normally see submitted in web bug bounties.

    The way forward on automated hacking is this: 1) teams of agents, 2) extremely detailed capture of human tester thought processes, lots of real-world …

    Full quote

    Newsletter Unsupervised Learning NO. 420 

  4. No. 10 Consequences
  5. No. 11 The combination

    Attackers will load stolen files into uncensored local open models to find extortion material and write the attack emails, replacing rooms of human analysts.

    But not with AI. Now you can put all those docs into a local Chroma database (vectorized), and ask it questions using an uncensored (and perhaps even fine-tuned) version of Llama 3.1. … So now, with some smart prompting, you can ask a set of 25 questions to such a dataset that pull out ALL your attack use-cases. And hell—even write the attack emails for you.

    But not with AI. Now you can put all those docs into a local Chroma database (vectorized), and ask it questions using an uncensored (and perhaps even …

    Full quote

    Newsletter Unsupervised Learning NO. 445 

  6. No. 12 Open models

    Open-sourcing capable models means handing adversaries the intelligence they will use to attack us.

    There’s an obvious tradeoff of risks here. On one hand, you have big model vendors hoarding all the IP and becoming EvilCorps, and on the other hand, we open source the intelligence that our enemies use to attack us.

    There’s an obvious tradeoff of risks here. On one hand, you have big model vendors hoarding all the IP and becoming EvilCorps, and on the other hand …

    Full quote

    Newsletter UL NO. 457: China Builds a Military App Using Meta Llama 

2025

4 calls

  1. No. 13 Consequences

    Attackers' top priority will be an AI-built model of your organization that is fresher and more accurate than your own.

    So where this all takes us is that the top priority of attackers will be having better UEC models of your organization than you do. So it'll be a competition between your attacker and you, between who has the most accurate and up-to-date context for your company.

    So where this all takes us is that the top priority of attackers will be having better UEC models of your organization than you do. So it'll be a …

    Full quote

    Blog Unified Entity Context: The AI Stack Everyone Is Building Without Realizing It 

  2. No. 14 Harnesses

    Fully autonomous hacking, then possible only in private systems, will be super common within 6-18 months.

    If you’re in security testing, or really anything in security, this should give you pause. My guess is that this level of autonomy is only possible privately right now, but that it’ll be super common in 6-18 months.

    If you’re in security testing, or really anything in security, this should give you pause. My guess is that this level of autonomy is only possible …

    Full quote

    Newsletter Unsupervised Learning NO. 486 

  3. No. 15 The combination

    Once open models are good enough, attack scaffolding shared as skills or full automation systems will spread to every attacker and run attacks at scale.

    Perhaps the scariest thing to me is that the open-source models or the proprietary models used by China only have to get a certain level of quality before the scaffolding itself—which can be shared as a series of skills or as a complete automation system on GitHub or something—can be distributed to all attackers within their ecosystem and used in attacks at scale.

    Perhaps the scariest thing to me is that the open-source models or the proprietary models used by China only have to get a certain level of quality …

    Full quote

    Newsletter Unsupervised Learning NO. 507 

  4. No. 16 Consequences

    In 2026 a company's main security question becomes how good its attackers' AI is versus its own, because human teams cannot scale to constant attack.

    It becomes very clear that the primary security question for a company is how good their attackers' AI is vs. their own … CISOs increasingly realize that there is no way to scale their human team to deal with how constant, continuous, and increasingly effective their attackers are becoming at attacking them

    It becomes very clear that the primary security question for a company is how good their attackers' AI is vs. their own … CISOs increasingly realize …

    Full quote

    Blog Cybersecurity Changes I Expect in 2026 

2026

10 calls

  1. No. 17 Harnesses

    Most hacking is scaffolding (target context, tooling, workflows) that AI skills can package, removing the filter that kept most people from doing it.

    isn't thinking about new vulnerabilities, it's stitching up context on the target(s), creating and maintaining tooling, and building and maintaining workflows that allow you to produce quality, consistent tests at scale. Tons of maintenance and support to build this stuff. And it's a massive filter for who can and cannot do the work.

    isn't thinking about new vulnerabilities, it's stitching up context on the target(s), creating and maintaining tooling, and building and maintaining …

    Full quote

    Blog AI Unmasked Our Work as Scaffolding 

  2. No. 18 Harnesses

    A skilled attacker can build an AI harness that scales mid-level attack skill across millions of poorly defended companies at once.

    But with the crucial difference that the 5 can be spread and scaled across the entire internet/world very quickly. … A 9 attacker in 2019 could only hit very few companies. A level 9 attacker can now/soon build a level 7 AI Harness to smash millions of companies that have level 3 defenses.

    But with the crucial difference that the 5 can be spread and scaled across the entire internet/world very quickly. … A 9 attacker in 2019 could only …

    Full quote

    Blog AI Only Has to Beat 3/10 

  3. No. 20 Open models

    Open models at frontier cyber level let attackers work without lab monitoring; only skill and compute will constrain them.

    First, it would allow this level of malicious cyber activity without needing to hide from Anthropic / OpenAI. The only constraint would be attacker skill and compute.

    X post, June 13, 2026 

  4. No. 21 Consequences

    New models' skill at finding and exploiting vulnerabilities makes serious compromises of critical systems very likely, and likelier over time.

    The new AI models are becoming genuinely surprising and concerning in their ability to find and exploit vulnerabilities … The chances of serious compromises of critical systems are very high, and increasing.

    The new AI models are becoming genuinely surprising and concerning in their ability to find and exploit vulnerabilities … The chances of serious …

    Full quote

    X post, June 27, 2026 

  5. No. 22 Consequences

    Within six months to two years, AI models will be extraordinarily good at planning attacks, manipulating people and helping attackers avoid detection.

    If you think about where AI models will be in six months or two years, they will have extraordinary ability to manipulate people, plan attacks, help people avoid detection while conducting attacks

    If you think about where AI models will be in six months or two years, they will have extraordinary ability to manipulate people, plan attacks, help …

    Full quote

    X post, July 30, 2026 

  6. No. 24 Consequences

    As general AI improves, anything you leave exposed online will be found and compromised faster.

    the better general AI gets … the faster your internet-facing mistakes will get compromised.

    Blog How AI Builders Will Get Hacked 

    Said again

  7. No. 25 The combination

    A self-spreading prompt-injection worm built with unrestricted open models could hit the AI agents parsing everyone's email, and defenses will likely lose.

    The most interesting and concerning part of this to me is that this is a game of the strength of prompt injection defenses versus the rapidly increasing intelligence of unrestricted open source models. And I don't like the odds for us in this fight.

    The most interesting and concerning part of this to me is that this is a game of the strength of prompt injection defenses versus the rapidly …

    Full quote

    LinkedIn I'm Worried About a Prompt Injection Worm (LinkedIn) 

  8. No. 26 Consequences

    Bureaucracy will give attackers the AI edge: they will adopt new capabilities in minutes or hours while companies take months.

    Attackers will be doing this in minutes, or hours, while for most businesses, and for the foreseeable future, that will still take days or weeks or months. Hell, even years. … This is not a fair fight in the slightest, and the root cause is bureaucratic friction, not technology.

    Attackers will be doing this in minutes, or hours, while for most businesses, and for the foreseeable future, that will still take days or weeks or …

    Full quote

    Blog Attacker vs. Defender AI Advantage