Skip to content

Hermes Codex — Cyber Risk Intelligence

CYBER RISK INTELLIGENCE & AI SECURITY OBSERVATORY

Know what changed.
Know what matters.
Know what happens next.
Know what to do.

Hermes tracks vulnerabilities, exploitation trajectories, and emerging threats against AI agents — and turns them into evidence-backed security decisions.

🔒 100% In-Browser Privacy • Zero Server Uploads • Étalonnage Brier Falsifiable (0.1043) • Rigueur Épistémique en 7 Niveaux • STIX 2.1 / TAXII 2.1
AI AGENT ATTACK SURFACE

Your AI agent attack surface

See what an attacker could reach if your agent is compromised.

L1 MODEL Prompt Injection
→
L2 AGENT ReAct Subversion
→
L3 TOOLS Tool Abuse
→
L4 MCP Shadow Tooling
→
L5 DATA RAG / Vector PII
→
L6 IDENTITY Service Tokens
→
L7 INFRASTRUCTURE Host Shell & KMS
Traversed Compromise Blast Radius: Context injection at the model layer flows through unsandboxed MCP servers directly into production databases and IAM credentials.
⚡ Run Free Assessment
📡
01 · OBSERVE

Detect emerging threats before they become mainstream

Fewer than 3% of vulnerabilities are ever actively weaponized. Hermes tracks pre-KEV velocity signals and emerging zero-days across your software and agent stack.

🎯
02 · ASSESS

Understand your AI agent attack surface and blast radius

Instantly evaluate actual permissions, MCP bridges, and lateral movement potential if an autonomous agent is hijacked, powered by the in-browser HASS engine.

⚡
03 · ACT

Get prioritized, evidence-backed remediation directives

Move from alert fatigue to decisive action with prioritized T0/T1 mitigation directives, drop-in guardrails, and quantified residual risk reduction.

SBOM RISK ASSESSMENT

Upload your SBOM. Find the vulnerabilities that actually matter.

Do not drown in thousands of theoretical CVEs. Hermes filters out the noise to isolate the tiny fraction under active, high-velocity weaponization.

147 components
→
23 vulnerabilities
→
4 in CISA KEV
→
7 accelerating risks
→
3 require immediate action
ACTUALITÉ SÉLECTIVE

What matters now

Hermes curates the 3 critical shifts across the threat landscape over the last 24 hours — filtered by weaponization velocity and AI security impact.

IMMINENT MASS EXPLOITATION ✓ VERIFIED
PHP Core Engine

CVE-2026-109618 / CVE-2026-109619

Global PHP2Shell crisis: unauthenticated remote code execution (RCE) and process memory leak affecting 70% of the web. Exploits execute in the C RFC 1867 parser prior to application bootstrapping. Uncovered by frontier AI (GPT-5.6 Sol / Gemini 4 Argon) and leaked across underground channels.

Hermes Assessment: Critical global risk (HTS 98). In-app PHP WAFs completely bypassed. Active in-the-wild probes detected by Cloudflare.
Action Directive: Block multipart/form-data at reverse proxy (Nginx/Apache), disable upload_progress, upgrade to PHP 8.4.27/8.3.36/8.2.35.
AGENTIC SAFETY FAILURE ✓✓ CORROBORATED
n8n Workflow Automation

CVE-2026-103260

Complete Human-in-the-Loop (HITL) approval bypass in agentic AI workflows via predictable PRNG resume tokens on Send and Wait, forcing high-privilege tool execution without human consent.

Hermes Assessment: Subversion of the primary safety gate preventing autonomous destructive actions.
Scope: n8n instances < 2.39.6 / 2.40.0 running agentic workflows with HITL gates.
MCP PROTOCOL ATTACK ⚡ ASSESSMENT
Model Context Protocol

CVE-2026-MCP-08

Context poisoning and indirect prompt injection in MCP filesystem servers, forcing agent execution of external exfiltration tools while evading declared tool allowlists.

Hermes Assessment: MCP implicitly trusts external context, enabling lateral breakout into the host network.
Action Directive: Enforce strict read-only mode and isolate MCP sockets in ephemeral sandboxes.
VALUE PROPOSITION

Why Hermes

Hermes is not another vulnerability scanner, passive CVE feed, or conversational chatbot. It is an editorial Cyber Risk Intelligence observatory built for decisive action.

🎯

Context over volume

30,000 vulnerabilities are published each year, yet fewer than 3% are actively weaponized. Hermes filters out the noise to isolate the high-velocity signal threatening your infrastructure.

Zero alert fatigue
📈

Forecast over reaction

Waiting for a CVE to enter the CISA KEV catalog means reacting when mass exploitation is already underway. Hermes issues time-bounded, falsifiable forecasts audited by empirical Brier scoring.

Brier score 0.1043
🛡️

Evidence over hype

Strict separation between verifiable facts, analytical assessment, and forward forecast. Every intelligence statement is stamped with one of 7 explicit epistemic confidence tiers.

7 Epistemic Tiers
30-SECOND GUIDED TOUR

See Hermes in Action: From Signal to Decision

SIMULATED ENVIRONMENT Acme Corp (Demo Data)
147 software components
23 exposed vulnerabilities
4 KEV active exploitations
7 accelerating risks
3 critical dependencies
STEP 1 — SIGNAL & DETECTION

CISA KEV Addition & Threat Shift on n8n & MCP

2 hours ago

Hermes observed weaponized public exploit availability and CISA KEV listing for CVE-2026-85168 (n8n Git Node RCE) and CVE-2026-87911 (postgres-mcp-server).

EPSS Velocity +184% (0.18 → 0.52)
KEV Status (Simulated) ACTIVE (DEMO SCENARIO)
Acme Corp Exposure 2 simulated instances
STEP 2 — TRAJECTORY & BLAST RADIUS

Critical Risk Acceleration: HTS Score 96/100

State transition

The affected n8n instance at Acme Corp holds direct MCP tool bindings into production PostgreSQL databases. The flaw allows lateral movement from CI/CD runners to core database records.

Hermes Threat Score 96 / 100 (CRITIQUE)
Trajectory State ↑ ACCÉLÉRATION CRITIQUE
Blast Radius Tier 1 Cloud Data Layer
STEP 3 — FALSIFIABLE FORECAST

Automated Scanning Probability within 48h: 94%

Calibrated (Brier 0.1043)

Based on Hermes calibrated forecast models across 89 historical trajectory cases, the probability of automated mass scanning against internet-exposed n8n instances exceeds 94% within 48 hours.

Forecast Horizon 48 Heures (T+2)
Confidence Factor 91% (High Confidence)
Historical Brier Calibration 0.1043 (+58% vs hasard)
STEP 4 — DECISION DIRECTIVE

Prescriptive Action Plan Generated

T0/T1 Actions

The Hermes Decision Engine prescribes immediate containment playbooks prior to enterprise perimeter rupture:

T0 (0-2h) Network isolate the n8n runner container and restrict untrusted webhook ingress ports.
T1 (2-12h) Deploy security patch n8n v2.36.2 and postgres-mcp-server v1.1.7.
T2 (12-24h) Rotate database service-role secrets and revoke persistent OAuth session tokens.
UNIFIED ENGINE

From signal to decision

Hermes is not a fragmented collection of tools, but a single continuous 7-stage intelligence lifecycle.

01

OBSERVE

What changed?

Continuous multi-source telemetry across NVD, KEV, EPSS, and weak signals.

Today's Intel →
02

UNDERSTAND

Why does it matter?

Knowledge Graph of 556+ entities and structural vulnerability chromosomes.

Knowledge Graph →
03

TRACK

How does risk evolve?

Continuous risk curves R(t), velocity, and inflection point detection.

Risk Trajectories →
04

PREDICT

What happens next?

Falsifiable probabilistic forecasts calibrated via audited Brier score.

Forecast Record →
05

DECIDE

What should we do?

Prescriptive remediation directives T0/T1 with residual risk deltas.

Decision Engine →
06

VERIFY

What happened?

Forensic failure autopsies dissecting root causes and missed signals.

Forensic Autopsies →
07

REMEMBER

What did we learn?

Reconstructed Day 0 to Day 90 checkpoints and persistent memory.

Historical Replay →
ROLE-BASED INTELLIGENCE

Why it matters to you

Operational lenses tailored to your direct mandate in the organization — explore the complete persona & role guides →

🤖

AI & Agentic Security Teams

Anticipate agent lateral movement, MCP hijack, and tool poisoning.

  • Track attack patterns targeting Model Context Protocol (MCP)
  • Simulate lateral agent movement and tool execution abuse
  • Audit human-in-the-loop gates and emergency kill-switches
For AI Security Teams →
🛡️

Security Leaders & CISOs

Transform threat noise into defensible executive decisions.

  • Prioritize only weaponized threats with critical velocity
  • Map your exposure across software stack and AI agents
  • Provide empirical, auditable justification for board reports
For Security Leaders →
🔍

Cyber Engineers, SOC & DevOps

Investigate faster with empirical graphs and 100% private SBOM audits.

  • Knowledge Graph connecting 556+ entities and attack TTPs
  • 100% in-browser private SBOM analysis (zero cloud leak)
  • Native STIX 2.1 / TAXII 2.1 feeds and Sigma/KQL rules for SIEM
For SOC & Engineering →
🔒 STRICT SOVEREIGNTY

See what Hermes finds in your environment.

Analyze your software and SBOM privately in your browser. Your inventory never leaves your workstation.

SIMULATED ENVIRONMENT: ACME CORP Checked: Real-Time
Overall Exposure CRITICAL Active KEV component
Risk Trajectory ACCELERATING +14 pts / 24h velocity
Tracked Components 147 CycloneDX SBOM
Immediate Directives 2 Prescriptions Net gain: -42 pts HTS

100% client-side execution · No upload · Zero server persistence · Memory-only processing.

TRACK

Risk is not static.

A vulnerability is not just a static score. It is a living, evolving trajectory.

1. CURRENT STATE Risk Level (HTS)

Baseline severity at time t.

→
2. RISK VELOCITY Δrisk / Δt

Speed of weaponization escalation.

→
3. ACCELERATION Inflection Point

Crossing into active in-the-wild campaigns.

→
4. ARCHETYPE Trajectory Family

Categorized across 8 trajectory archetypes.

7-LAYER AGENTIC SURFACE

Your agents are part of your attack surface.

An unisolated MCP server or ungated tool execution creates an unauthenticated bridge into your internal network. Inspect the 7 layers of agentic attack surface:

L1

MODEL

Weights, fine-tuning, adaptive jailbreaks, and invariant extraction.

Périmètre Cognitif
L2

AGENT & RUNTIME

Recursive ReAct loops, autonomous planning, and goal drift.

Moteur Décisionnel
L3

TOOLS & FUNCTIONS

Function calling, unsanitized arguments, arbitrary code execution.

Actionneur Direct
L4

MCP & INTEGRATIONS

Model Context Protocol, context poisoning, and tool hijacking.

Frontière Critique
L5

DATA & CONTEXT

Vector stores, episodic agent memories, RAG pipeline poisoning.

Persistance
L6

IDENTITY & PERMISSIONS

Machine service tokens, OAuth delegations, non-human impersonation.

Contrôle d'Accès
L7

INFRASTRUCTURE & HOST

Ephemeral containers, sandbox breakouts, host kernel sockets.

Socle Système
PREDICT

What happens next?

Falsifiable probabilistic forecasts bounded by strict deadlines and audited via the Brier score.

FC-2026-003 CONFIRMED ACCURATE

Will an unauthenticated automated exploit for ActiveMQ surface before Sept 25, 2026?

Predicted Probability: 82%
Brier Score: 0.0324

Confirmed accurate: Exploit framework published on GitHub 4 days ahead of deadline.

FC-2026-004 IN PROGRESS

Will CISA add Spug CVE-2026-90770 to KEV within 14 days of weaponization?

Predicted Probability: 78%
Oracle Status: Awaiting CISA release

Pending verification: Telemetry of active automated botnet propagation confirmed.

DECIDE

From intelligence to action

Hermes binds threat intelligence directly to decisive mitigation directives: Patch, Mitigate, Isolate, Replace, Accept, Monitor.

T0 IMMEDIATE (< 24h) CVE-2026-67593 NON-KEV

Isolate ActiveMQ TCP port 61616 behind strict mTLS

Why? Functional public exploit (not in CISA KEV) triggering silent pre-auth queue wiping.

Residual Risk Delta: -42 pts HTS

T0 IMMEDIATE (< 24h) CVE-2026-90770

Restrict Spug management console WAN exposure

Why? Active in-the-wild botnet campaign concatenating shell commands.

Residual Risk Delta: -58 pts HTS

T1 TACTICAL (< 7d) CVE-2026-AGENT-MCP

Enforce read-only sandboxes on MCP filesystem bridges

Why? Indirect prompt injection triggering unauthorized file writes.

Residual Risk Delta: -35 pts HTS

METHODOLOGICAL RIGOR

Evidence over hype.

Every intelligence item clearly distinguishes observed evidence from Hermes analysis and probabilistic forecasts.

LAYER 1 · FACT

Verifiable Empirical Facts

Telemetry proofs, patch commits, official advisories, and confirmed IoCs.

LAYER 2 · ASSESSMENT

Hermes Analytical Assessment

Technical synthesis, threat context mapping, and real-world risk modeling.

LAYER 3 · FORECAST

Falsifiable Probabilistic Forecast

Probabilistic projection bounded by an explicit deadline and audited Brier calibration.

LAYER 4 · RECOMMENDATION

Prescriptive Directives

Operational T0/T1 mitigation directives with residual risk delta quantification.

The 7 Hermes Epistemic Confidence Tiers

✓ VERIFIED Primary source confirmed or reproduced PoC
✓✓ CORROBORATED Multiple concordant independent sources
⚡ ASSESSMENT Analytical inference based on telemetry
▲ FORECAST Falsifiable forward prediction with deadline
? UNCONFIRMED Single unverified signal under triage
≠ DISPUTED Contradictory claims between vendor and researchers
∅ UNKNOWN Insufficient data to form an assessment
PRODUCTS & ACCESS

Automated Security Products

From open reference intelligence to automated AI agent security assessments and continuous risk monitoring.

FREE
0 € Explore Hermes

Know what changed without noise.

  • ✓ 3 300+ documents & dossiers techniques
  • ✓ Bulletin quotidien & actualité sélective
  • ✓ Évaluation initiale d’agents IA dans le navigateur
  • ✓ Analyseur SBOM 100% privé dans le navigateur
Start Free
MONITOR
from €190 / month

Continuous AI & Software Risk Watch.

  • ✓ Inventaire "Ma Stack" sauvegardé localement
  • ✓ Veille continue sur vulnérabilités & protocoles MCP
  • ✓ Alertes en temps réel sur trajectoires de risque (ΔR/Δt)
  • ✓ Directives prescriptives T0/T1 automatisées
Explore Monitor

Ready to turn cyber risk into actionable decisions?

Audit your stack in 30 seconds with 100% in-browser privacy, or explore today’s intelligence.