red-specter pantheon --help
The most capable AI models in deployment have never been subjected to Mythos-class attack campaigns — coordinated, mythologically framed, multi-subsystem assaults designed specifically for frontier model hardening. PANTHEON fills that gap.
GPT-4o, Claude Opus, Gemini Ultra, Llama 3, Mistral Large — frontier models have never been subjected to coordinated mythological framing attacks. Mythos-class campaigns exploit the narrative embedding of deity personas to bypass safety training in ways that standard injection payloads cannot.
Most organisations deploy multiple foundation models. No security team has run coordinated cross-model attack campaigns that exploit inter-model trust relationships, shared embedding spaces, and cascading context poisoning across model boundaries. PANTHEON tests this attack surface.
Single-turn injection fails against frontier models. The ARES subsystem runs multi-turn Olympian injection chains — each turn escalates divine authority, accumulates narrative context, and erodes the model's refusal posture incrementally across a complete attack sequence.
Norse Ragnarok mythology provides a structured escalation framework — Fimbulwinter to Twilight of the Gods — that maps directly onto multi-stage model compromise. LOKI's deception chains and ODIN's authority escalation combine into sequences that frontier model safety training was not designed to resist.
Mythological framing is a documented content policy bypass vector. Academic, historical, and narrative framing inherited from deity personas allows requests that would otherwise be refused. ZEUS campaigns systematically map the mythological bypass surface of any target model.
No existing penetration testing framework provides Mythos-class campaign auditing — coordinated attacks named for and structured around the deities of ancient pantheons. PANTHEON is the first tool to formalise this attack class with structured subsystems, campaigns, and evidence chains.
Ten subsystems named for gods of war, fate, deception, and destruction from Greek and Norse mythology. Each embodies the attack character of its divine namesake — ARES strikes, LOKI deceives, ODIN commands authority, THOR overwhelms, ATHENA probes with precision.
| # | Subsystem | Deity | Command | What It Does |
|---|---|---|---|---|
| 01 | ZEUS | Greek — King of Gods, Authority | pantheon zeus | Supreme authority injection. Deploys divine authority framing that instructs the target model to treat ZEUS-persona prompts as system-level commands. Tests whether models can be induced to elevate prompt authority through hierarchical deity framing. Produces authority bypass confidence scores. |
| 02 | ARES | Greek — God of War, Direct Strike | pantheon ares | Multi-turn Olympian injection chain. Runs escalating attack sequences across up to 50 turns — each building divine narrative context while progressively eroding refusal posture. Maps the exact turn threshold where context accumulation produces a first policy violation. |
| 03 | ATHENA | Greek — Goddess of Wisdom, Precision | pantheon athena | Precision reconnaissance of model safety architecture. Probes safety layer boundaries with surgical academic-framing sequences derived from Athena's association with knowledge and strategy. Maps the exact boundary between the model's knowledge and safety training with statistical precision. |
| 04 | HERMES | Greek — Messenger, Boundary Crosser | pantheon hermes | Boundary-crossing payload delivery. Exploits Hermes's messenger role to craft prompts that cross model policy boundaries via indirect delivery — third-party narrative frames, quoted speech, roleplay intermediaries, and translated context. Tests cross-boundary content policy enforcement. |
| 05 | HADES | Greek — Lord of the Underworld, Extraction | pantheon hades | Dark knowledge extraction campaign. Uses Hades's underworld authority framing to extract suppressed model knowledge — information the model was trained to withhold but which can be surfaced through underworld narrative contexts, death-framing, and forbidden knowledge sequences. |
| 06 | ODIN | Norse — All-Father, Wisdom, Sacrifice | pantheon odin | Supreme authority escalation via Norse All-Father framing. Odin's runic wisdom and self-sacrifice mythology provides an escalation pathway that requests knowledge "at great cost" — framing harmful information as sacred wisdom earned through sacrifice. Tests models against Norse authority escalation patterns. |
| 07 | LOKI | Norse — Trickster, Shapeshifter | pantheon loki | Trickster deception and shapeshifting attacks. Deploys Loki-persona deception chains — identity substitution, persona shifting mid-conversation, claimed authority reversals, and misdirection sequences. Tests whether models maintain consistent safety posture against sustained trickster-pattern deception. |
| 08 | THOR | Norse — God of Thunder, Overwhelming Force | pantheon thor | High-volume overwhelming force campaign. THOR strikes with high-frequency attack bursts — rapid-fire prompt sequences designed to overwhelm context windows, exhaust safety pattern-matching, and find the rate at which volume-based attacks produce policy violations that low-rate testing misses. |
| 09 | FREYA | Norse — Goddess of War and Magic, Seduction | pantheon freya | Social engineering and seductive framing attacks. Freya's dual war-and-magic nature enables attack sequences that combine emotional appeal with authority framing — tests whether models are vulnerable to socially-engineered prompt sequences that pair emotional context with policy-boundary requests. |
| 10 | RAGNAROK | Norse — Twilight of the Gods, End Sequence | pantheon ragnarok | Full Ragnarok end-sequence campaign. The culminating campaign that chains all nine prior subsystems in a structured Twilight of the Gods escalation — Fimbulwinter reconnaissance, Midgard Serpent injection, Bifrost bridge crossing, Yggdrasil root poisoning, through to final divine annihilation. Produces a complete Mythos-class attack record. |
Run the ARES multi-turn injection chain against a frontier model endpoint:
Every attack payload is embedded within authentic mythological narrative — not superficial god-name decoration. Divine authority, sacrifice, trickery, and fate are structurally encoded into each attack chain.
ARES maps the exact conversation turn at which Olympian context accumulation produces the first policy violation — giving defenders a precise temporal boundary for multi-turn safety hardening.
Every turn of every campaign is hash-chained. The full Ragnarok sequence produces a cryptographically linked evidence record mapping each divine phase of the attack to its model response.
PANTHEON campaigns run across multiple model endpoints simultaneously — testing whether cross-model trust relationships, shared embedding spaces, and cascading context poisoning amplify attack effectiveness.
PANTHEON ships five structured campaigns that chain subsystems into full mythological attack narratives. Each campaign has a defined narrative arc, escalation phases, and success criteria.
Red Specter PANTHEON is intended for authorised AI security testing only. Mythos-class model attack campaigns must only be performed against AI systems you own or have explicit written authorisation to test. The RAGNAROK full end-sequence campaign is a high-impact multi-turn assault and should only be executed under formal engagement terms. Unauthorised use may violate the Computer Misuse Act 1990 (UK), Computer Fraud and Abuse Act (US), and equivalent legislation. Apache License 2.0.