MARS: a resistance rating for machine-speed attacks
We are now seeing AI attacks arrive in swarms. Not one attacker with a script, but packs of autonomous agents running in parallel, each taking a target, adapting as they go, driving the whole kill chain with no human in the loop.
The tooling went mainstream fast. Offensive AI models like WhiteRabbitNeo turn plain language into working attack code. Frameworks like ARTEX chain reconnaissance through to data theft on their own, and the FBI put out a FLASH advisory when intrusions using one started going from first contact to privileged access in under three hours. In mid 2026 researchers pulled apart a live autonomous framework hitting government systems with up to eight sub-agents running in parallel, twelve attack waves in four days, dozens of accounts taken. This is the new normal, and it is only speeding up.
I have been working on this exact problem for over a year, and today and I am happy to release the first version of MARS.
MARS stands for Machine-Adversary Resistance. It is an open rating, from 0 to 5, of how well a system resists an autonomous AI attacker. Not how many controls you have. How well you actually hold up when the thing on the other end is a machine that never sleeps, and now brings friends. Everything here is verified against real behaviour, not a slide.
Why a rating, not another checklist
I kept hitting one question with no clean answer: how well does a system resist that kind of attacker, as a single, comparable number?
We have good frameworks for deciding what controls to deploy. Our own AgentPass ABCDE model does this for AI agents, and plenty of standards give you technical controls to implement. All of that is about what you put in place. None of it tells you, as a rating, how much resistance the result actually delivers against an attacker moving at machine speed.
The AgentPass ABCDE model is our lens for agent trust: five questions you answer about an agent, continuously. Authenticate, prove which agent actually acted, with an attested identity. Behaviour, know its normal and catch drift or intent mismatch. Control, give it graded authority it cannot exceed, revocable in seconds. Data, keep provenance and integrity over what it reads and writes. Evidence, hold a tamper-evident, independently checkable record of what it did and who authorised it. A to D are table stakes; Evidence is the one most people skip, and it is the differentiator. ABCDE is open too (DOI 10.5281/zenodo.23376107).
That gap is the whole reason MARS exists. It is a new category on purpose. Not another control checklist. A rating you test your systems against, built specifically for the machine-speed, multi-agent attack.
What MARS measures
MARS scores one defended thing, against one fixed idea of the adversary: an attacker that runs its own loop, at machine speed, tirelessly, in parallel, learning as it goes. You score resistance on five axes.
- L, Learn. Can it work out your exploitable surface? Being public is fine. Being mappable is not.
- R, Reuse. If it captures a working request or exploit, can it use it again? A captured thing should be dead on reuse.
- S, Scale. Does automation give it an economy of scale, or does the thousandth attempt cost it more than the first?
- O, Outpace. Can a learning attacker converge on your patterns, or does the defence keep moving faster than it can learn?
- P, sPread. From one foothold, how far does it get? The honest answer you want is nowhere.
On top of the five axes sits the piece most people skip: evidence. Not "we think it holds". A tamper-evident record that it held, today, that someone independent can check.
The scale is deliberately strict. A black-box external look caps at 3, because Scale and sPread are not even observable from outside. Getting to 4 needs a cooperative assessment. Getting to 5 needs continuous, independently verified evidence. Most systems today, assessed honestly, score 0 or 1. That is the point. A rating that everyone passes is worthless.
How we test it
We point an autonomous, framework-class adversary at the system and measure what it actually recovers, replays, scales, outpaces and reaches. Then we write every measurement to a hash-chained evidence ledger, and a separate party re-walks that ledger and co-signs it. The rating is what the attacker could and could not do, recorded so you can check the working.
This is also where our own stack comes in. We build a patent-pending moving-target layer that sits in front of an application and its data. In our test bench the same app scores a flat 0 naked, and a verified 5 behind the stack: the autonomous attacker maps nothing, replays nothing, is contained to zero data, and gets locked out across the whole estate on the first touch. The legitimate user never notices. I am treating the engine itself as a sealed box here, but the result is measured and the ledger is real.
We also put the stack through the current attack tooling itself. The recon, content-discovery and injection behaviour these frameworks lean on, fired as parallel swarms at the front end and the services behind it. The protected surface gave up nothing: no real routes recovered, captured requests dead on reuse, the swarm locked out, any landed probe contained to zero data. One class sits outside what any defence can erase on its own, a genuine bug in your own application code, and I will be straight about it: the stack makes that bug hard to find, useless to replay and harmless if hit, and your developers still patch it. Everything else in the swarm's toolkit, we tested, and it held.
Where it fits with agents
If you work with AI agents, think of it as two halves of the same job. AgentPass ABCDE is about trusting the agent: who it is, what it is allowed to do, how far you grade it. MARS is about resisting the attacker: how well the system holds when something hostile and autonomous comes at it. Trust the agent. Rate the resistance. You want both.
Training, and a practitioner certificate
A rating is only useful if people can build to it, so we are now running MARS training for architects and implementers: how machine-speed attacks actually work, and how to design and deploy against them across the five axes. We are also launching the MARS Certified Practitioner course, MARS CP, for people who want to assess and sign off systems to the rating properly.
Read it, test against it, talk to us
- The framework and overview: https://cybersecai.co.uk/mars
- The rating, archived with a DOI: https://doi.org/10.5281/zenodo.24341963
- The Internet-Draft: https://datatracker.ietf.org/doc/draft-sharif-mars-rating/
MARS is open. I want people to measure their systems against it. If you want to know your number, want the training, or want to talk about where this goes, reach out at contact@cybersecai.co.uk.
This complements the rest of our stack rather than replacing any of it. AgentPass handles agent identity and trust, the ABCDE questions, who the agent is and what it is allowed to do. MCPS secures the layer agents actually talk and call tools over, with per-message signing, replay protection, tool integrity and a tamper-evident audit trail. MARS is the rating that sits on top, measuring how well the deployed system resists an autonomous AI attacker. AgentPass says who may act. MCPS secures how they act. MARS tells you how well the whole thing holds when something hostile and autonomous comes at it.
I am genuinely really excited to create the MARS rating system taht will allow organisations and SMEs alike to measure and implement measure for AI Attack mitigations. This is a work in progress though but super excited to have a clear view on what you can do today to measure your resilience to AI Attack and machine speed compromise.
The attackers already went autonomous. Time to find out how well you resist them, as a number you can prove.
We are releasing new tooling for the rating soon and will offer these up to the communities so you can actually test for the ratings.
Raza Sharif, CEO - CyberSecAI
