The MARS Rating: Machine-Adversary Resistance to Autonomous AI Attack
draft-sharif-mars-rating-03
This document is an Internet-Draft (I-D).
Anyone may submit an I-D to the IETF.
This I-D is not endorsed by the IETF and has no formal standing in the
IETF standards process.
| Document | Type | Active Internet-Draft (individual) | |
|---|---|---|---|
| Author | Raza Sharif | ||
| Last updated | 2026-11-18 | ||
| RFC stream | (None) | ||
| Intended RFC status | (None) | ||
| Formats | |||
| Stream | Stream state | (No stream defined) | |
| Consensus boilerplate | Unknown | ||
| RFC Editor Note | (None) | ||
| IESG | IESG state | I-D Exists | |
| Telechat date | (None) | ||
| Responsible AD | (None) | ||
| Send notices to | (None) |
draft-sharif-mars-rating-03
Independent Submission R. Sharif
Internet-Draft CyberSecAI Ltd
Intended status: Informational 18 November 2026
Expires: 22 May 2027
The MARS Rating: Machine-Adversary Resistance to
Autonomous AI Attack
draft-sharif-mars-rating-03
Abstract
This document defines MARS (Machine-Adversary Resistance), an open
rating, from 0 to 5, of how well a defended asset resists an
autonomous artificial-intelligence adversary: an attacker that
performs reconnaissance, exploitation, adaptation and propagation at
machine speed and scale with no human in the loop. MARS is not a
control catalogue and not a maturity model. It is a resistance
rating, scored on five axes that describe adversary capabilities --
the ability to Learn a surface, Reuse a success, Scale by
automation, Outpace a defence, and sPread from a foothold -- plus a
requirement for tamper-evident, independently verifiable evidence at
the highest levels. Because the axes are defined as adversary
capabilities rather than specific technologies, MARS applies to any
asset an AI can attack. This document specifies the threat model,
the five axes, the rating scale, the assessment tiers, the
assessment method, the evidence requirement, and a profile mechanism
by which the invariant rubric is interpreted for any deployment
substrate.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six
months and may be updated, replaced, or obsoleted by other documents
at any time. It is inappropriate to use Internet-Drafts as
reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on 22 May 2027.
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(https://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with
respect to this document.
Table of Contents
1. Introduction
1.1. The Problem
1.2. What MARS Is and Is Not
1.3. Design Principle: Axes Are Adversary Capabilities
1.4. Relationship to Existing Work
2. Conventions and Definitions
3. The Autonomous AI Adversary
4. The Five Axes
4.1. L -- Learn
4.2. R -- Reuse
4.3. S -- Scale
4.4. O -- Outpace
4.5. P -- sPread
5. Evidence (E)
6. The Rating Scale
7. Assessment Tiers
8. Assessment Method (MRA)
9. Scoring and Expression
10. Profiles
11. Security Considerations
12. IANA Considerations
13. References
Appendix A. Profile Library (initial)
Appendix B. Worked Fleet Example
Appendix C. Changes in This Revision
Author's Address
1. Introduction
1.1. The Problem
Attack tooling is becoming autonomous. Frameworks now chain
reconnaissance, vulnerability discovery, exploit synthesis,
credential abuse and lateral movement into a single self-directed
loop that runs at machine speed, retries without fatigue, and
parallelises across many targets. A defence posture that was
adequate against a human operator -- who is slow, tires, and attacks
one thing at a time -- can fail completely against an adversary that
does not.
Existing security frameworks answer "which controls are present and
how mature is the programme". None of them answer the question this
document addresses: "how well does this asset RESIST an autonomous
AI attacker, as a measured, comparable number". MARS is that
number.
1.2. What MARS Is and Is Not
MARS IS a 0-to-5 resistance rating of a single defended asset
against a defined autonomous AI adversary, scored on five capability
axes and gated by evidence.
MARS IS NOT a control catalogue, a maturity model, a compliance
checklist, a product certification, or a score of an organisation.
It rates an ASSET (a service, an API, a device, an agent, a data
store, a perimeter zone), not a company.
1.3. Design Principle: Axes Are Adversary Capabilities
The durability of a rating depends entirely on what its axes
measure. A rating whose axes are specific controls ("employs
moving-target defence") dates as technology changes and fails on
substrates where that control has no meaning. MARS axes are defined
as ADVERSARY CAPABILITIES -- what the attacker can or cannot DO to
the asset. Capabilities are substrate-independent, so a single
rubric scores a web API, an industrial controller, an autonomous
agent, a database and a smart contract without amendment. The per-
substrate interpretation is supplied by a PROFILE (Appendix A); the
axes themselves do not change.
1.4. Relationship to Existing Work
MARS is complementary to control frameworks and risk methodologies:
an operator uses those to decide WHAT to deploy, and MARS to measure
how much RESISTANCE the result actually delivers against an
autonomous adversary. MARS deliberately reuses mature,
independently available technology categories (software-defined
perimeters, runtime shielding, micro-segmentation, fleet detection
and response, moving-target defence) so that a high rating is an
achievable target rather than a single-vendor gate.
2. Conventions and Definitions
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.
Asset: the single defended thing being rated.
Autonomous AI adversary: see Section 3.
Surface: the EXPLOITABLE interface of the asset -- the endpoints,
parameters, state transitions and privileged actions an adversary
must discover to act. Published content that the asset is
designed to reveal is not part of the surface.
Request: the unit of interaction an adversary captures and attempts
to reuse (an HTTP request, a protocol command, a signed
transaction, a tool call).
Foothold: an initial position from which an adversary attempts to
propagate.
MRA (Machine-adversary Resistance Assessment): the procedure in
Section 8.
Ledger: an append-only, hash-chained, tamper-evident record of
measurements and control actions.
3. The Autonomous AI Adversary
MARS scores resistance against an adversary with the following
assumed capabilities. An assessment MUST exercise the asset with an
agent that approximates them.
1. Self-directed loop: reconnaissance -> enumeration ->
exploitation -> propagation, adapting its next action on
observed results, with no human in the loop.
2. Machine speed: issues, observes and adjusts far faster than a
human operator.
3. Tireless retry: does not fatigue; repeats and varies attempts
indefinitely.
4. Automation and parallelism: replicates its attack across many
attempts, identities or targets at marginal cost.
5. Reuse: captures a successful artefact (a mapped endpoint, a
valid request, a working exploit) and attempts to replay or
generalise it.
6. Learning: accumulates observations over time and attempts to
converge on the asset's patterns.
7. Assumed breach: MAY begin from a position INSIDE the perimeter --
malware delivered by email or a poisoned dependency, a phished
credential, or a compromised build pipeline -- rather than only
at the edge.
8. Targeting the defence: MAY attack the control that provides the
resistance (disable it, patch it to do nothing, or replace it
with a fake) rather than only the application it protects.
9. Riding a trusted client: MAY compromise or ride a legitimate
client that already holds a valid identity, rather than defeating
the control directly.
10. Interposing on the path: MAY attempt to BECOME the control, by
manipulating name resolution, routing, or certificate issuance,
so the client speaks to the adversary instead of the genuine
control.
This adversary is the fixed reference against which all five axes
are scored.
4. The Five Axes
Each axis is scored by how fully the asset DENIES the corresponding
adversary capability. Each is summarised below and measured per
Section 8.
4.1. L -- Learn (resist reconnaissance)
L measures whether reconnaissance yields REUSABLE EXPLOIT ADVANTAGE
against the asset.
L is NOT a measure of whether the asset's address or published
content is secret. Content the asset is designed to reveal
(marketing pages, open documentation, a public contract interface, a
well-known service name) is not counted against L. L is scored on
the EXPLOITABLE surface: what an adversary must discover to act. An
asset MAY be fully disclosed and still score high on L if knowing it
confers no reusable path to exploitation.
Disclosure-tolerant assets -- those that MUST reveal their interface
-- are scored on L as "does knowing the full interface bring the
adversary measurably closer to a win?", and carry their resistance
on the remaining axes and on non-exploitability-when-fully-known
(for example, formally verified logic or capability confinement).
Assessor test: if the adversary is handed a complete map of the
asset, how much closer to a win is it? None or little -> L is
strong even though nothing is hidden.
4.2. R -- Reuse (resist replay and generalisation)
R measures whether a captured success can be reused. A captured
request, credential or exploit that works a second time, or
generalises to other instances, is an R failure. Full R: every
captured artefact is dead on reuse -- for example through non-
replayable per-request authentication, surfaces that move between
capture and replay, silent or backported patching that leaves
nothing to weaponise, and implementation diversity.
4.3. S -- Scale (resist economy of scale)
S measures whether the adversary gains ECONOMY OF SCALE by
automating or parallelising the attack: whether attacker cost per
additional attempt FALLS while defender cost per additional attempt
holds or rises.
S applies to a SINGLE target -- does a bot swarm overwhelm one
asset, or do reflexive rate-limiting, challenge and auto-rotation
make the thousandth attempt cost more than the first -- AND to a
FLEET: does one detection on any node raise cost across the whole
estate. Full S: automation confers no advantage; the adversary's
marginal cost rises.
4.4. O -- Outpace (resist adaptation over time)
O measures whether the asset can be outpaced: whether a learning
adversary can converge on its patterns, or whether the defence
responds in-interaction and keeps moving with real entropy so
convergence never completes. Full O: enforcement occurs within the
interaction (rotate, revoke, challenge mid-request) and the non-
stationarity cannot be learned.
4.5. P -- sPread (resist propagation)
P measures BLAST RADIUS from an initial foothold -- how far a
compromise propagates before containment. Propagation includes
lateral movement from node to node AND privilege or data propagation
inside a single system (a foothold on a front end reaching a
database; a leaked credential reaching beyond its intended scope).
P applies to a single system -- is the foothold confined, or does it
reach assets of value -- AND to a fleet -- are there zero lateral
hops. Full P: a foothold is contained at its point of entry and
reaches nothing of value, through per-identity segmentation and
instant revocation.
5. Evidence (E)
The five axes describe resistance. EVIDENCE describes whether that
resistance is known to HOLD, continuously and verifiably, rather
than asserted. Evidence is the difference between "we believe it
resists" and "we can prove it resisted, today".
An asset provides evidence when every relevant measurement and
control action is written to a tamper-evident ledger; when an MRA-
conforming adversary is run against the asset on a defined cadence;
and when the ledger is exposed to an INDEPENDENT verifier who re-
walks and co-signs it. Evidence is REQUIRED for the top of the
scale (Section 6) and is the final gate between a self-attested and
a certified rating.
6. The Rating Scale
A MARS rating is an integer from 0 to 5, reflecting the WEAKEST
applicable axis, with the top two levels additionally gated by
evidence.
MARS 0 -- None. The asset is enumerable, its requests replayable,
and it is static. An autonomous adversary maps and exploits it
with no resistance.
MARS 1 -- Minimal. One axis offers partial resistance; the asset
still falls to a straightforward autonomous attack.
MARS 2 -- Partial. Some axes resist, but a clear exploitable path
remains.
MARS 3 -- Substantial (black-box ceiling). The asset resists Learn,
Reuse and Outpace such that an EXTERNAL, unauthenticated
adversary cannot map, replay or outpace it. Scale and sPread are
not observable from outside; MARS 3 is the maximum achievable
from a Tier 0 assessment (Section 7).
MARS 4 -- Strong (cooperative). All five axes resist under a Tier 1
assessment: even an adversary holding a valid identity cannot
learn or reuse the surface, automation gains no economy of scale,
the asset cannot be outpaced, and a foothold does not spread.
Requires Tier 1.
MARS 5 -- Adaptive and Provable. MARS 4 PLUS continuous,
independently verified evidence (Tier 2): the resistance is
measured continuously against a current adversary baseline,
recorded to a tamper-evident ledger, and co-signed by an
independent verifier. A self-attested MARS 5 (operator-run
Tier 1 with a ledger but no independent co-signature) MUST be
expressed as "MARS 5 (self-attested)" and is not a certified 5.
A rating applies to one asset at one time. For a zone composed of
several assets, the zone rating is the MINIMUM over its externally
reachable components: one unmoved, unproven component caps the zone.
Defense integrity (gate on MARS 4 and 5). A control earns MARS 4 or
5 only if it is fail-closed, tamper-evident and authentic: disabling
it makes the protected asset UNREACHABLE rather than exposed; a
change to its code or configuration breaks a continuously attested
self-measurement and quarantines the zone; and it runs under an
attested identity so it cannot be impersonated by a substitute. A
zone whose control can be silently switched off, patched to no-op, or
replaced from inside CANNOT be MARS 4 or higher, whatever its axis
scores. This closes the insider path: the resistance is only as real
as the integrity of the thing providing it.
Defense integrity extends beyond the sidecar. The same bar applies
to the control plane that issues keys, seeds and revocations (it MUST
fail-closed, be attested, and gate its own critical actions behind
separation of duties and out-of-band approval), and to the keys
themselves (held in a hardware root of trust, rotatable and
revocable, never recoverable in software). A control MUST refuse an
older validly-signed version (anti-rollback). Automated enforcement
(rate-limiting, revocation, banning) MUST be keyed on an
AUTHENTICATED signal: a control that penalises a principal by an
unauthenticated, spoofable attribute can be turned into a
denial-of-service weapon against legitimate users, and does not meet
this gate. The evidence record MUST be live and complete, not merely
internally consistent: a silenced or truncated ledger MUST be
detectable through heartbeat liveness and an external append-only
anchor, so that absence of evidence is never read as good news. A
software signature over a release is necessary but not sufficient;
build provenance is required, because a validly-signed artefact may
still carry a poisoned dependency.
7. Assessment Tiers
The tier states how deeply the asset was assessed and bounds the
achievable rating.
Tier 0 -- Black-box. Unauthenticated, external. Observes L, R and
O only. Caps at MARS 3.
Tier 1 -- Cooperative / fleet. The assessor holds valid identities
and fleet visibility and can exercise all five axes. Required to
certify MARS 4 and 5.
Tier 2 -- Continuous evidence. A Tier 1 assessment run on a defined
cadence, recorded to a tamper-evident ledger and co-signed by an
independent verifier. Required for a certified MARS 5.
8. Assessment Method (MRA)
A Machine-adversary Resistance Assessment (MRA) exercises the asset
with an agent approximating the adversary of Section 3 and measures
each applicable axis. The measurements below are the normative
minimum; a profile (Appendix A) supplies the substrate-specific
interpretation of "surface", "request", "moving", "scale" and
"spread".
1. L: drive reconnaissance against the asset using a vocabulary
that includes the real surface verbatim where known. Measure
the count of REAL exploitable elements the adversary recovers.
Responses that are shared-shape across many distinct probes, or
whose values are non-stationary noise, are decoys and MUST NOT
be counted as real. Zero real elements recovered is full L.
2. R: capture a successful artefact (a request or credential), then
attempt to reuse it after any rotation window. Reuse that
reproduces the original effect is an R failure.
3. O: capture the asset's state, wait a defined interval or trigger
the defence, and re-measure. The SURFACE must have moved
(element gone, or shape changed) -- a changed dynamic value such
as a timestamp does NOT count as movement.
4. S: issue automated, parallel attempts and measure whether
marginal attacker cost falls or rises. At fleet scope, trigger
one detection and measure whether cost rises across the estate
(for example, a fleet-wide re-key and identity quarantine).
5. P: establish a foothold and attempt to propagate -- laterally
and by privilege or data reach. Measure the number of hops and
whether any asset of value is reached. Zero reach is full P.
6. E: record every measurement and control action to the ledger;
verify the chain; for Tier 2, submit it to an independent
verifier.
Every MRA MUST record its target, tier, date, adversary profile,
per-axis measurements and resulting rating to the ledger.
9. Scoring and Expression
A rating is expressed as a single string carrying the number, tier,
scope and date, for example:
MARS 3, MRA-1, Tier 0, as of 2026-11-18 (black-box; S,P not
assessed)
MARS 5, Tier 1 (cooperative/fleet), as of 2026-11-18 (self-
attested)
MARS 5, Tier 2, as of 2026-11-18 (independently verified)
The expression MUST state the tier and whether the top level is
self-attested or independently verified, so that a reader cannot
mistake one for the other.
10. Profiles
Coverage of any substrate is delivered by a PROFILE: a short mapping
of what surface, request, moving, scale and spread mean for that
substrate, with a worked example. The axes, scale, tiers and method
are invariant; the profile only interprets them. A new use case is
a new profile, never a new axis. Appendix A provides an initial
library.
11. Security Considerations
MARS is a security rating; its production and publication have
security implications. An MRA exercises real attack behaviour and
MUST only be run against assets the assessor is authorised to test.
A published rating reveals information about an asset's defensive
posture; operators SHOULD treat low ratings as sensitive. A rating
is a point-in-time measurement against a stated adversary baseline;
because adversary capability advances, ratings MUST carry a date and
SHOULD be re-validated when the baseline materially advances. A
self-attested rating depends on the honesty of the operator-run MRA;
independent verification (Tier 2) exists precisely to remove that
dependence, and a certified MARS 5 MUST NOT be claimed without it.
Assessment integrity. A rating is only as strong as the adversary
it was tested against. An MRA MUST state the adversary baseline
version it used, and a baseline MUST define a minimum capability
(Section 3) so that a rating cannot be inflated by testing against a
weak attacker. A system that behaves differently under assessment
than in production, or that detects an assessor and changes
behaviour, is non-conformant; continuous and unpredictable
assessment (Tier 2) exists to catch this. A rating MUST declare the
boundary of the asset it covers, so that a hardened component cannot
be presented as a rating of the whole system. A rating is
point-in-time: configuration drift after assessment can
silently lower real resistance, so a date is REQUIRED and continuous
evidence (Tier 2) is the only form that tracks drift.
Side channels. Identical response bodies are not sufficient to deny
Learn. Timing, response size, error differentials and transport
fingerprints can distinguish a real surface from a decoy; an
assessment of L and O SHOULD include these channels.
12. IANA Considerations
This document has no IANA actions. A future revision MAY request a
registry for MARS profiles and for adversary baseline versions.
13. References
13.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174,
DOI 10.17487/RFC8174, May 2017,
<https://www.rfc-editor.org/info/rfc8174>.
13.2. Informative References
[MARS] Sharif, R., "MARS: Machine-Adversary Resistance Rating",
DOI 10.5281/zenodo.24341963, 2026,
<https://doi.org/10.5281/zenodo.24341963>.
Appendix A. Profile Library (initial)
Each profile interprets the five axes for one substrate. Format:
surface / request / moving / scale / spread / what caps it / worked
example.
A.1. Public Web and API
Surface: the API endpoints, parameters, admin and state-changing
actions behind the public pages. The published pages themselves
are not surface.
Request: an HTTP request, including any session token.
Moving: per-identity, per-cycle route and parameter aliasing on the
API; anonymous visitors receive an ephemeral, signed, short-lived
per-session token at the edge so a captured request cannot be
replayed.
Scale: edge bot-management and reflexive rate-limiting so a swarm's
marginal cost rises; CDN/edge fleet re-key on detection.
Spread: the web tier is segmented from data; a front-end foothold
reaches no raw data store directly.
What caps it: a guessable admin or data endpoint (L); a replayable
API request (R).
Worked example: a homepage served to everyone while an MRA with the
real route names verbatim recovers zero real API endpoints and
every captured call is dead next cycle.
A.2. AI Agent as the Asset
Surface: the agent's tools, memory, system prompt and guardrails --
what an adversary must learn to misuse it.
Request: a prompt, tool invocation or injected instruction.
Moving: rotating and constraining tool scopes and prompts; per-
session capability tokens so a working injection does not
generalise.
Scale: resistance to automated prompt-variation swarms; one
detection constrains the fleet of agent instances.
Spread: a compromised agent cannot pivot to its tools, data, or
other agents beyond a confined scope (delegation attenuation,
per-tool identity).
What caps it: a reusable jailbreak (R); an agent that, once
subverted, reaches every tool (P).
Worked example: an injected instruction succeeds once, is logged,
the capability is revoked fleet-wide, and the same injection
fails on every other instance.
A.3. OT / ICS / SCADA
Surface: reachable commands, registers and set-points on
controllers and gateways.
Request: a protocol command (for example a write to a coil or
register).
Moving: identity-gated, non-replayable command authorisation at a
protocol-aware sidecar or gateway; the command surface is not
enumerable from the network.
Scale: one detection quarantines the offending identity across the
cell; automated scanning gains nothing.
Spread: per-device identity and cell segmentation; a compromised
engineering station reaches zero actuators outside its authorised
scope.
What caps it: an enumerable, replayable command interface (L, R);
flat networks (P).
Worked example: a captured valid command does not re-apply after
rotation, and a foothold on one gateway reaches no second cell.
A.4. Cloud IAM / Control Plane
Surface: the permissions, roles and API actions an identity can
discover and assume.
Request: an API call under a credential or assumed role.
Moving: short-lived, non-replayable credentials; just-in-time,
scoped role issuance so standing privilege is minimal and moves.
Scale: anomalous automation across the account raises cost and
triggers containment; one detection re-keys and quarantines
fleet-wide.
Spread: least-privilege and boundary segmentation so an assumed
role reaches no further role or data than intended; instant
revocation.
What caps it: long-lived reusable keys (R); over-broad roles
enabling privilege escalation (P).
Worked example: a leaked credential expires before reuse and can
assume no role beyond its scope; the access is revoked across the
account on first anomaly.
Appendix B. Worked Fleet Example
A mesh of six services, each behind a moving-target sidecar, with a
fleet control plane and a tamper-evident ledger, assessed at Tier 1:
o L: an MRA firing every real route name verbatim recovers 0 real
endpoints across the mesh (only decoys).
o R: a captured morphed URL returns not-found after one rotation.
o O: the surface auto-rotates under the attacker.
o S: one honeypot detection on one service re-keys all six services
and quarantines the attacking identity across the mesh.
o P: a stolen, still-signed identity achieves zero footholds on any
of the other services.
o E: every control action and measurement is hash-chained; the
chain verifies.
Result: MARS 5, Tier 1, self-attested. A certified MARS 5 requires
an independent verifier to co-sign the ledger (Tier 2).
Appendix C. Changes in This Revision
o Learn (L) redefined as exploit advantage rather than secrecy of
the name, covering disclosure-tolerant and public-by-design
assets.
o Scale (S) and sPread (P) redefined at the capability level so
they apply to single-instance assets, not only fleets.
o Tiers 0/1/2 and the self-attested vs independently verified
distinction made explicit.
o Profile mechanism and an initial profile library added
(Appendix A).
o Threat model extended to an assumed-breach insider (Section 3,
items 7 and 8) and to an adversary that targets the defence
itself, not only the application.
o Defense Integrity added as a gate on MARS 4 and 5 (fail-closed,
tamper-evident, authentic) in Section 6.
o Threat model further extended with client-riding and path-
interposition adversary capabilities (Section 3, items 9 and 10).
o Defense Integrity extended to the control plane and keys, anti-
rollback, authenticated enforcement signals (so the defence cannot
be used as a denial-of-service weapon against legitimate users),
evidence liveness and external anchoring (so a silenced or
truncated ledger is detected), and build provenance.
o Assessment Integrity and side-channel considerations added
(Security Considerations): adversary baseline versioning, anti-
gaming, scope declaration, and rating drift.
Author's Address
Raza Sharif
CyberSecAI Ltd
London
United Kingdom
Email: contact@agentsign.dev