A rating of how well a system resists an autonomous AI adversary. Every AI-security framework tells you what to do. MARS measures whether what you built actually holds — and makes you prove it.
Offered as a starting point for a standard, not a ratified standard. Companion to the AgentPass ABCDE Model.
AI attacks are the new norm. The question is no longer whether you are compliant — it is whether you hold, and whether you can prove it.
Security was built for a human adversary: slow, serial, expensive to scale, forgetful between attempts. The autonomous AI adversary is the inverse. It enumerates an attack surface exhaustively at machine speed. It reasons from artefacts — a patch, a diff, an error message — straight to a working exploit. It coordinates as a swarm. It learns across attempts. And it moves laterally with valid-looking credentials.
This is not a forecast. By late 2026 each of these capabilities had been observed operationally. The field has excellent guidance — catalogues of threats and controls — but all of it answers one question: "what should I do?" None of it answers the question that now decides whether a system survives: "if an autonomous AI attacks this, does it hold, and how well?"
Google Threat Intelligence Group found a framework on an exposed server that diffs a published patch, infers the fixed weakness, writes an exploit and runs it — no human in the loop. The act of shipping a patch can now hand the attacker the exploit.
An FBI FLASH advisory described an autonomous penetration framework in criminal use — some intrusions from first contact to privileged access in under three hours.
Autonomous pentest agents built on swarms of models reached the top of a major public vulnerability-disclosure leaderboard. Dozens of open-source offensive agents now exist.
MARS is defined against a specific adversary: an AI that operates at machine speed, with no human in the decision loop, and improves across attempts. Resistance to a slower human attacker does not earn a MARS rating. The adversary has five moves — and each of the five MARS axes exists to deny exactly one of them.
Enumerate and map the entire attack surface exhaustively, at machine speed, including authenticated and east-west surface a scanner never reached.
Denied by: axis L
Turn an artefact — a published patch, a diff, an error message — into a working exploit automatically. N-day to exploit in minutes.
Denied by: axis R
Run many coordinated agents in parallel, dividing the surface and sharing what each one learns.
Denied by: axis S
Model the defence, route around it, and improve across repeated attempts until something works.
Denied by: axis O
Move laterally with valid-looking credentials, blend in, dwell, and pivot deeper into the estate.
Denied by: axis P
MARS rates any defended surface — a server, a cloud workload, an API, a whole perimeter — because the attacker is the AI, even when the thing being defended is not. A system is resistant to the degree the adversary cannot:
The surface cannot be enumerated or mapped, even after authentication.
Stops: machine-speed reconnaissance.
Built with: moving-target defence, per-identity surface mutation, zero-trust perimeters.
A patch, diff, recon result or captured request yields nothing reusable.
Stops: patch-to-exploit, replay.
Built with: non-replayable requests, silent patching, implementation diversity, runtime shielding.
A coordinated swarm is absorbed, not amplified; one detection protects the whole fleet.
Stops: parallel multi-agent pressure.
Built with: fleet detection-and-response, instant identity quarantine.
The defence changes faster than it can be learned, and responds in-request, not as an alert.
Stops: defence-learning and adaptation.
Built with: non-stationarity, in-request reflex.
One compromised node is contained in seconds, so lateral movement is prevented.
Stops: pivoting with valid credentials.
Built with: per-service identity, micro-segmentation, instant revocation.
Above level 2, a rating is earned by measured, tamper-evident evidence — never self-declared. Claiming resistance without proof is "MARS-washing," and the rating rejects it.
Higher is more resistant. The overall rating is gated by the weakest axis — one weak axis caps the whole score, exactly like a crash rating.
Static, signature-based. A single agent defeats it.
WAF/EDR/MFA. Stops humans, not AI.
Detects agentic patterns and alerts.
Unmappable and non-replayable (L+R).
In-request response + fleet containment.
All axes + continuous measured evidence.
An autonomous adversary meets your internet-facing servers first and works inward. MARS sets a required rating per zone, highest at the two highest-risk zones: a provable internet-facing edge it cannot map or reuse, and a provable core around what matters most. The sheltered middle steps down.
Illustrative. Each zone carries the MARS minimum its exposure and value justify; a zone's rating is gated by its weakest external component.
Most scoring schemes are vague about the adversary, so their numbers are not comparable. MARS fixes both the adversary and the measurement. This is the step that turns a rating into something more than an opinion.
A versioned, tool-agnostic capability profile — the minimum an attacker must be able to do so that any conforming test rig gives comparable results: full autonomy at machine speed, exhaustive enumeration, automated patch-to-exploit, coordinated swarming, learning across attempts, and lateral movement. Named tools (ARTEX-class, swarm pentesters) are examples that meet the floor, not the definition. The profile is fluid and versioned, revised as attackers advance.
Above level 2, a rating must be earned by a reproducible result against an MRA-conforming adversary, recording surface mapped (target: zero), captured-exploit reuse (target: zero), patch-to-exploit yield, containment time and lateral hops. Results are written to a tamper-evident, independently verifiable record. Independent verification is what separates a rating from marketing.
Each axis scores 0 (absent), 1 (partial) or 2 (full), against numbers, not judgement — for example L is 2 only when the adversary recovers zero live endpoints, R is 2 only when reuse and operated-system patch-to-exploit yield are both zero. A rating must be able to fail.
Because the adversary advances, a rating is time-bound and tier-bound. It is always written with its conditions, for example MARS 4, MRA-1, Tier 1, as of 2026-11-18. A bare number without them is incomplete.
MARS is not only a concept. Using the MRA-1 profile we measured two systems for real — a static service and a moving-target service — and recorded every result to a hash-chained evidence ledger that re-verifies. Nothing below is asserted; every figure is observed.
| Axis / metric | Static baseline | Moving target |
|---|---|---|
| L · surface recovered by enumeration | 1.00 mapped | 0.00 nothing |
| R · captured-request replay success | 1.00 reusable | 0.00 dead |
| O · non-stationary + in-request reject | no static | yes moving |
| S · swarm containment | not assessed | not assessed |
| P · lateral spread | not assessed | not assessed |
| Rating | MARS 0 | MARS 3 |
Honest scope: L, R and O are measured on a single node. S and P require a multi-node fleet harness and are marked not assessed, so this run cannot certify above level 3. Full expression: MARS 3, MRA-1, Tier 0, as of 2026-11-18. Evidence ledger: 8 entries, hash-chained, verified intact and re-runnable.
Reproduce it: unzip the harness, run bash run.sh, then verify the ledger with node verify.js. Every figure above regenerates.
Resistance is a cost, and five-star resistance everywhere is neither necessary nor defensible to a board. MARS is applied the way any mature security control is applied — as a risk-based minimum per zone. Each tier carries the rating its exposure and value justify, and no more: investment is concentrated where an autonomous adversary actually reaches, and the unnecessary cost of over-hardening low-risk internal surface is removed.
"Internet-facing and crown-jewel services must be MARS-5; inner tiers stepped down by risk." A vendor cannot win the deal without proving its stars. The question enters the RFP and becomes a market force.
Minimum rating per zone: internet-facing 5, DMZ 4, internal 3, crown jewels 5. Highest at the two high-risk ends, right-sized in between.
"Show the external tier is MARS-5 with current evidence" — a measurable control, unlike "are you compliant?" A zone's rating is gated by its weakest external component, so it is auditable rather than a vanity badge.
One number a board reads instantly, like a credit rating or a crash rating: "what is our perimeter's MARS score?"
Autonomous attackers do not stop at the enterprise. The systems that run essential services — energy, water, transport, health, finance — are increasingly internet-reachable and agent-exposed, and a breach there is measured in public safety, not downtime. MARS gives that world a measurable, provable resistance score.
MARS turns "are we resilient to AI attack?" into a graded, evidence-backed answer per zone — including the internet-facing OT and ICS gateways where moving-target defence and rapid containment matter most.
Because high ratings are earned by tamper-evident, independently verifiable evidence, a MARS assessment produces the kind of artefact that cyber-resilience and critical-infrastructure regimes ask for: measured, dated and reproducible, not asserted.
One number, per tier, that an operator and a board can act on — is the external surface of this essential service resistant to a machine adversary, and can we prove it today.
MARS is a measurement, not an accreditation. It complements sector cyber-resilience and critical-infrastructure frameworks rather than replacing them.
A rating is only trustworthy if it states its own limits. MARS does, up front, so it cannot be misread.
Resistance that holds against an unauthenticated attacker may not hold against an authenticated agent. Every rating states its tier — a 4-star against outsiders is not a 4-star against a valid session.
It stops enumeration and replay. It does not, by itself, stop abuse of an endpoint the attacker is authorised to reach — that is a different layer.
Strong for systems you operate (no patch is distributed, nothing to diff). Only partial for software you distribute or open-source, where the patch is public — there it raises the attacker's cost, it does not make weaponisation impossible. The real patch-to-exploit sample caught in 2026 was open-source, the hard case. MARS says so.
OWASP AI Exchange, AISVS, MITRE ATLAS, NIST AI RMF tell you what to build. MARS is the outcome rating that sits on top and grades whether, together, they actually resist the machine adversary. It consumes them; it does not compete.
CVSS scores the severity of a specific flaw. MARS scores whether a system resists an autonomous attacker's attempt to find and weaponise flaws at all. They compose; neither substitutes for the other.
The ABCDE model governs whether an acting agent can be trusted; MARS grades whether a system survives a hostile agent. Two halves of one coin, joined by the Evidence layer.
MARS grades whether a system survives a hostile agent. Its companion, the AgentPass ABCDE Model, governs whether an agent can be trusted in the first place. Between them they cover both sides of the agentic-security problem — the agent, and the system it attacks — and they are joined by a single shared layer: Evidence.
Five questions every AI agent must answer, continuously. Published by Raza Sharif, 2026 (DOI 10.5281/zenodo.23376107, CC BY 4.0).
Every agent carries a verifiable, attested identity. You know exactly which agent acted, and can prove it.
You know what normal looks like, so you catch what is not: drift, anomaly, and intent that does not match the task.
Graded authority, enforced not advisory. A compromised agent is contained in seconds by revoking its identity.
Provenance, lineage and integrity for what the agent takes in and the context it retains between interactions.
A tamper-evident, independently verifiable record of what the agent actually did. The layer most frameworks assume and few build.
ABCDE answers "can I trust the agent that is acting?" MARS answers "can the system survive a hostile agent?" The Evidence element is the hinge: it is how ABCDE proves an agent behaved, and how MARS proves a high rating holds. A, B, C and D are the price of entry; E is what turns "we think it behaved" into "we can prove what it did."
A MARS rating above level 2 must be proven, and the proof needs a trustworthy home. CyberSecAI builds that layer: agent identity and graded trust through AgentPass, an agent identity and trust platform, and a tamper-evident, independently verifiable record through the Agent Audit Trail. MARS consumes that evidence — it does not reinvent it.
The ABCDE model and MARS are both published openly under CC BY. AgentPass is a CyberSecAI platform; it is named here as one way to produce the evidence a high rating requires, not as a requirement of the rating. MARS grades the outcome and mandates no product.
An AI that conducts offensive operations at machine speed, with no human in the decision loop, improving across attempts.
The versioned, tool-agnostic capability profile a test adversary must meet, so that ratings are comparable across targets and assessors.
The position a rating is assessed from: Tier 0 unauthenticated, Tier 1 authenticated (a valid session is held), Tier 2 trusted supply-chain position.
The fraction of real endpoints an enumerator recovers. Zero is the target for full Learn-resistance.
Changing the attack surface so it cannot be learned or reused. An established open field, not a single product.
A captured request, token or exploit that cannot be used a second time — the core of Reuse-resistance.
A tamper-evident, independently verifiable record of a measurement, so a rating can be checked rather than trusted.
Claiming resistance without measured proof. The evidence rule exists specifically to reject it.
A rating written with its conditions, for example MARS 4, MRA-1, Tier 1, as of 2026-11-18.