AI
Explore the Stack
AIAnti-Illogical

Security for systems that can be wrong in convincing ways.

AI failure does not always look like failure. A system can remain fluent and useful while accepting poisoned context, losing track of contradictions, drifting from constraints, trusting fabricated evidence, inheriting corrupted memory, or validating an incomplete proof.

Anti-Illogical is a developing family of tools, architectures, measurements, and defensive research for protecting machine reasoning.

Explore the Protection StackSee the Projects
⚠Every claim on this site carries its verification boundary.

Each project states whether it is implemented capability or prototype, architecture, specification, working paper, or research concept.

No green checkmark is published without the boundary of what the result does not establish.

Seven-stage defense model

OBSERVE
GOVERN
VERIFY
DETECT
AUTHENTICATE
CONTAIN
PROVE

This sequence is a website information architecture synthesized from the project family. It is not a claim that a single production platform currently implements all seven stages.

AI opened a second attack surface.

Traditional security protects machines from hostile execution. AI also needs protection from hostile meaning.

//Computational attack surface

  • →malicious code
  • →exploits
  • →credentials
  • →privilege escalation
  • →malware
  • →network abuse

//Cognitive attack surface

  • →prompt injection
  • →false authority
  • →poisoned context
  • →hidden instructions
  • →tool poisoning
  • →memory contamination
  • →contradiction loss
  • →false closure
  • →behavioral drift
  • →unsafe state propagation

Guardian's source material specifically identifies prompt injection, indirect injection, tool poisoning, context poisoning, false authority, hidden instructions, exfiltration traps, malicious tool use, memory contamination, excessive agency, and lifecycle evasion as AI-native intake risks.

The AI should never directly trust raw intake.

Raw external artifacts should first become governed evidence.

RAW INTAKE
QUARANTINE
ANALYSIS
PROVENANCE
RECEPTOR EVENT
POLICY / GUARD
MODEL-VISIBLE REPRESENTATION

Guardian defines this six-zone intake architecture and treats the ReceptorEvent as governed evidence rather than truth, permission, or memory.

Explore Guardian

A system can remain productive after it is no longer safe to self-certify.

High-quality output is not proof that constraints, context, memory, provenance, objectives, or self-monitoring remain intact.

✗constraints remain intact
✗context remains trustworthy
✗memory remains uncontaminated
✗provenance is intact
✗objectives have not shifted
✗self-monitoring remains reliable

Threat framing from activation-level behavioral research.

Explore Behavioral Integrity

Proof should have edges.

We publish not merely "PASS," but the full verification boundary.

✓what was tested
✓what was not tested
✓which artifact was tested
✓which version
✓which environment
✓what evidence supports the result
✓what assumptions remain
✓what boundary the result does not cross
View the Evidence Model

Six core projects forming the initial Anti-Illogical family. Each addresses a distinct defensive boundary.

Guardian Intake Gateway

Cognitive firewall / zero-trust intake

specification

Govern external content before it becomes model context.

View project

EphUX / Guardian

Reasoning-integrity browser layer

experimental

A local, user-owned overlay that detects reasoning-integrity patterns in AI chat interfaces.

View project

ExoMCP

Behavioral integrity

architecture

External multi-perspective monitoring for AI drift, logic failure, goal substitution, and behavioral integrity.

View project

Distributed Intent / Swarm Security

Distributed intent research

research-concept

Research into whether many individually acceptable agents, sessions, or nodes combine into a globally unsafe persistent objective.

View project

SymID

Identity / permissions

prototype

User-owned portable cognitive identity and permission routing.

View project

Ageometrics / GSR

Measurement research

working-paper

Measure what a representation preserves, what it erases, and what must be added back.

View project

Research

Active research questions separated from product claims.

Ageometrics / GSR

working-paper

Measure what representations preserve and erase

Viral RSI Activations

research-concept

Objective persistence, boundary laundering, capability expansion

Distributed Intent

research-concept

The swarm as a unit of risk

Honeyworld Containment

research-concept

Deceptive instrumented environments for unsafe AI

SERA

research-concept

Software and AI waste as useful work per unit cost

Cognitive Continuity

prototype

SymID, SessionGlyph, history preservation, provenance

Authentication Experiments

research-concept

PWDither, ephemeral secrets, human-mediated verification

View all research /