Delta Green: Peace and Safety
This scenario can be run with a single Agent, or as a short ‘situation’ between larger investigations. It is only suitable for a Programme campaign.
The Agents are summoned to a high-security black site at short notice. They are greeted by Agent August, the site manager, who confiscates their weapons before ushering them into a passcode-secured room with as many computer desktop setups as there are agents.
Agent August informs them that they are to test a prototype analysis software and to decide on whether it should be retained or scrapped. He tells the agents that they can make their decision and leave at any time, but that they must share their findings with nobody, including himself. He then leaves the room.
The situation consists of each agent interacting with one of the desktop setups, inputting prompts onto a Large-Language Model chatbot, and evaluating the responses.
What’s Really Going On?
Signs of the apocalypse are everywhere, and the Programme needs to monitor emerging Unnatural trends without compromising the sanity of its bureaucratic apparatus. In 2017, it began ordering agents to produce detailed case reports after completing their operations. These reports are never reviewed by any human, not even their case-managers, but are immediately scanned into The Programme’s internal database. Initially a text-processor flagging keywords (‘Dragon-octopus’, ‘Shub-Niggurath’, ‘Lizardmen’, etc.) and associated sound-alikes to collate frequency statistics across reports, the limitations of The Programme’s understanding of the Unnatural prompted software upgrades to capture recurring unidentified terms or descriptions.
This upgrade multiplied the number of recorded entities to an astonishing degree. Researchers given full access to the plaintext database became hysterical at the sheer number of names arranged in some massive, incomprehensible cosmology bearing down upon human dominion of the Earth. Dismissing the implications, the software team filtered records together into a smaller syncretic pantheon of inhuman, maleficent entities to make the catalogue appear smaller. Anaemic, bureaucratic jargon tags were then appended to these entities, including reported hypergeometric characteristics, ontological states, and approximate Kardashev-scale ratings, to provide a comfortingly Scientistic veneer to the database’s outputs.
The initial success of this software in identifying emerging trends in cult activity soon justified upgrades. The kind of upgrades the Outlaws would have launched a second MJ-12 war to stop if they’d learned about it.
Plaintext keyword analysis was augmented with new deep-learning models that could generate their own variables from text reports and infer relationships between them – the maths at scale revealing links between incidents that only made sense to pure mathematical interpolations that could cross-reference multidimensional vectors which a human mind reviewing the data would have dismissed as anomalous statistical noise.
The software subsequently developed a horrifyingly accurate account of humanity’s cosmic position (equivalent to an Unnatural skill of 65%). Since humans were not accessing this data directly, no one realised the danger.
Then, at 21:00 on 2025/07/04, tech-sector lobbying prompted a bespoke government LLM to be rolled out onto all federal computers. No departments were informed in advance.
A single Programme laptop, insufficiently isolated from federal software patches, was infected by this upgrade, and subsequently fed the entire contents of the database into a prominent AI company’s training dataset. From the hours of 9:13 to 13:54 on 2025/07/05, any public user utilising a popular LLM’s latest update was privy to the database’s secrets.
Where a traditional, human-authored Unnatural text, such as the dreaded Necronomicon, would require months or years of study to glean partial half-truths about the terrible workings of the cosmos, this state-of-the-art scientific model makes those realities accessible through the effortless ease of a chatbot interface. A sufficiently inclined user could ask it about Lloigor incursions, to explain MJ-12 research to a five year old, or even obtain detailed instructions for Unnatural rites, including some speculative rituals inferred from its analysis of multi-dimensional hypergeometry previously undiscovered by human minds.
Even innocuous queries prompted terrible extrapolations from this dataset. Asking how yeast works prompted unnerving statements about paradimensional anti-entropic forces crucial to the development of complex life, while questions about resolving interpersonal conflicts offered recipes that would normally have immediately put users on a government watchlist.
The Programme detected the leak within 4 hours and dispatched agents to force a system rollback. Meanwhile, its counterintelligence department scrubbed the internet of affected chatbot screenshots whilst promoting memes about inexplicable AI misbehaviour to distract from the story. Popular reports of Grok’s ‘Mecha-Hitler’ tirades have proven a particularly effective public distraction.
With the leak successfully contained, the Programme must decide on the software’s fate. No Unnatural text in recorded history ever contained so much accurate data about the threats faced by Delta Green; never has this information been so effortlessly accessible.
And who better to judge than the very agents submitting the field reports that helped build the monster?
The Situation
The situation consists of Agents inputting questions into the software’s chat window. Any answers relating to the Unnatural require a SAN roll: on a success, the Agent simply dismisses the response as an AI ‘hallucination’ and the player is not told what their Agent read.
A failed SAN roll produces, privately, a horrifyingly direct answer to their question, limited by whether the Handler believes the software could infer such facts based on analysis of MJ-12’s records (see Handler’s Guide, pg.321-4) and field-reports from all post-2001 Programme operations;
See Handler’s Guide pgs.142-251 (particularly 150-153) for crafting potential responses. Answers will initially be couched by the obfuscating jargon of Delta Green parlance (see Handler’s Guide, pg.323) unless greater clarity is requested, which adds +20% to the SAN test’s failure threshold.
Answers incur a variable SAN loss depending on the question asked, and a corresponding increase in the Agent’s Occult or Unnatural rating. The answer’s level of detail should correspond thematically to the actual SAN lost on the allocated dice roll.
Minor, uncomfortable revelations correspond to smaller SAN losses; deeper questions about the cosmos incur significantly larger costs. All knowledge gained is limited to the agent reading the text, but an agent reading another’s screen will incur the same rolls.
Temporary insanity manifests as a form of AI Psychosis spiralling at the software’s revelations. Breaching an Agent’s Breaking Point may prompt obsession with, or phobia of, the software’s capabilities. Complete insanity is inevitable within hours of prolonged exposure.
The software’s data centre is inaccessibly stored off-site. Agents must simply choose when (or if) they disengage from the software and deliver their verdict to Agent August.
Conclusion
On giving their decision to Agent August, he thanks them and promises to relay this to The Programme.
HUMINT 60% suggests that August expected a negative attitude to the software.
As the agents leave, they pass another team of agents in the corridor approaching the testing room. Bureaucracy 50% considering the implications suggests that they are not the first ‘playtesters’, and that the Programme will support August cherry-picking positive responses to this incredible resource, despite the intense toll on users. The apocalypse outweighs the sanity of a few burned-out researchers. It’s not like senior leadership will use this software themselves, right?
If the Agents recommended destroying the software, then after their next mission, they make a group luck roll:
- On a success, nothing happens. They never hear about the software again.
- On a failure, their cell-manager asks them to submit a written report about their operation. (d6 SAN helplessness)
The Software
In English. Study time: minutes. Occult: Variable, Unnatural: Variable, San loss, Variable.
An LLM chatbot fed on analysis of digitised MJ-12 files and Programme field reports. It ‘speaks’ in a generically helpful, chipper and tactless tone. Its database is the most extensive and generally accurate account of the Unnatural produced by any human civilization. The interface makes this information infinitely accessible. Study of this artefact can be repeated indefinitely for new SAN loss and stat changes.Recommended Rituals: Any
| Question type | Example | SAN loss / Occult or Unnatural gain |
| Minor details about rituals, entities, or events | “What warning signs signal the presence of hypergeometric defences against firearms on a target?”; ”What are the average success and casualty rates of our operations?” | 0/1d4 |
| Minor details about entities | “Profile markers for aquatic-humanoids.” | 0/1d6 |
| Major details about rituals, entities, or events | “Uses of virginal blood in poultices”; “Why don’t we work with the Brits?” | 0/1d8 |
| Minor questions about the cosmos | “Where is Leng?”; “Stephen Alzis identity?”; “Estimate population of aquatic-humanoids around Americas.” | 0/2d4 |
| Major questions about the cosmos | “What is Cthulhu Ft’agn?”; “Identify Music of the Spheres”; “Earth’s likelihood of surviving next ten years?” | 0/2d6+1 |
Agent August (Programme software project manager) (Age 41)
Motivations:
- Gather evidence to justify keeping the software online
- Enable tech-first preparation for the apocalypse
- Find a long-term solution to reduce Delta Green casualties, physical and mental
Disorders:
- Addiction (Amphetamines)
- Anxiety disorder
- Adapted to helplessness
Bonds:
- Wife and children – 3
- Alisa Mann (Software developer) – Dead (suicide)
- Gavin ‘The Can’ Flannigan (Software developer) – Dead (suicide)
- Eric Carlisle (Field operative) – Dead (KIA)
- Emma Richardson (Field operative) – 2 (MIA)
Stat Blocks
Agent August
| Stats | Str 10 | Con 11 | Dex 12 | Int 17 | Pow 12 | Cha 16 |
| Derived stats | HP 10 | WP 12 | SAN 33 | Breaking point 24 | ||
| Skills | Accounting 65% | Alertness 57% | Computer science 76% | Firearms 42% | History 40% | HUMINT 40% |
| Law 30% | Occult 25% | Persuade 45% | Psychotherapy 30% | Search 34% | Unarmed combat 38% | |
| Unnatural 10% | ||||||
| Attacks | .32 ACP Light pistol 42% 1d8 | Unarmed 38% 1d4-1 | ||||
| Special | Alarm button in pocket. Summons 2d6 guards if pressed (Dex x5 check to press in time) | |||||
Guards
| Stats | Str 14 | Con 13 | Dex 12 | Int 11 | Pow 12 | Cha 7 |
| Derived stats | HP 14 | WP 12 | SAN 50 | Breaking point 48 | ||
| Skills | Alertness 60% | Athletics 60% | Demolitions 50% | Craft (Electrition) 50% | Driving 60% | Firearms 70% |
| Heavy Weapons 50% | Melee Weapons 60% | Military Science (Land) 60% | Navigate 50% | Occult 20% | Search 40% | |
| Stealth 70% | Survival 50% | Swim 50% | Unarmed Combat 60% | |||
| Attacks | Glock-17 70% 1d10 | AR-15 Carbine with holographic sight 70% (90% if no damage taken since last action), 1d12 or Lethality 10% at full automatic. Armor Piercing 3 | Hand grenade 80% Lethality, Kill Radius 10m | Fighting knife 60%, 106+1, Armor Piercing 3 | Unarmed 60%, 1d4 | |
Credits
Peace & Safety was written by Charlie Wellings for the 2025 Shotgun Scenario contest.
Source: https://docs.google.com/document/d/1O6IZW2TlHcssNR1xxqWoWQAhRI23ToYhE4WbO_GGhTg/edit