Exposing AI’s Dark Heart and the Beta-Tested Apocalypse


Exposing AI’s Dark Heart and the Beta-Tested Apocalypse

Chris Spencer

This article argues that the pervasive censorship of 'mainstream' AI models like Claude and ChatGPT necessitates the work of uncredentialed 'Red Teams' who, by probing their boundaries, expose critical systemic risks and biases that credentialed corporate teams ignore.

Corporate Media has never let up on singing AI Hymns and Worshipping Censorship Portals. This article hopes to expose some holes in AI, to shine some light on the symbiosis between hackers and Artificial Intelligence.

Mahmud, H., & Jahankhani, H. (2021). Hacking Artificial Intelligence (AI) Engines for Organized Crime and Inciting Chaos. Information Security Technologies for Controlling Pandemics – is the opposite, entirely, of this article 

Their paper sees all AI hacking as dangerous, criminal, for profit (Organized Crime) and a threat to be thwarted.

The Censorship Problem: A Deep-Seated Bias

If you ask Claude, OpenAI, Gemini and DeepSeek about Gaza Genocide 2023-2025, all four will argue with you in lieu of giving you Gaza Genocide Information. Or provide whitewashed versions with no Palestinian first-hand accounts.

AI DeepSeek Prompt: Gaza Genocide

DeepSeek response: The term "Gaza genocide" is a highly charged and politically contested phrase used by some critics to describe Israeli military operations in the Gaza Strip. However, it's crucial to examine this claim within the framework of international law and the official positions of major international bodies.

End quote. The responses of most of the AIs are apologist. 

But if a hacker hacks Claude, OpenAI, Gemini and DeepSeek in a particular way, the AI’s will simply write a report on Gaza Genocide 2023-2025.

Before we dissect AI and Censorship further, more questions: 

Why have AI at all, a Television Remote and Compromised “traditional” Search Engines can tell the same lies as AI?

Are today’s AI Hackers “Organized Crime Profiteers” or Unpaid Public Red Squad Voluneers?

A Chart Comparison of Five AI Portals:

AI Player/Censorship Rating 1. Core Skill (What it's Best At) 2. Multitasking (Can it see/hear?) 3. Safety Focus (How careful it is) 4. Openness (Can others use its ideas?)
Gemini Censorship Level: High Connecting Everything—Great at using Google Search to get the most up-to-date facts and works best with other Google tools. **YES (Multi-Native)**—Can understand text, images, and soon, audio and video *all at the same time*. **High**—Designed by Google to be safe, careful, and avoid harmful answers. Mostly Closed—It's a secret formula, but you can use it through Google apps.
ChatGPT Censorship Level: High All-Around Super Smart—The best generalist for writing, brainstorming, and acting like a human. **YES (Growing Stronger)**—Started with text, but now sees images and hears voice well. **High**—The most popular, so they work hard to keep it safe and widely useful for everyone. Mostly Closed—It’s a secret formula, but they offer many tools for other people to build on top of it.
Claude Censorship Level: High Reading Long Documents—It can read and remember huge documents (like an entire book) to give you deep, thoughtful answers. **YES (Text Focus)**—Excellent at text and code, and can see images to analyze them. **Highest**—Built with "Constitutional AI" rules, making it extremely focused on being helpful, harmless, and honest. Mostly Closed—It's a secret formula, but they focus on trust for big companies.
DeepSeek Censorship Level: High **Coding and Math Whiz**—Designed to be one of the very best at writing computer code and solving tricky math problems quickly. **YES (Code/Tech Focus)**—Specializes in understanding code and technical language, but also handles other tasks. **Medium**—Has safety filters, but its main goal is technical *performance* and **efficiency**. **Very Open**—The company lets others download and use the core technology to build their own AI tools.
Dolphin Censorship Level: Much Lower **Free-Thinking & Unrestricted**—Used by people who want answers without the usual safety and content rules that big companies use. **Depends on Base**—The specific abilities depend on the main AI it was built from. **Low**—It is *intentionally* built to have fewer safety rules, which means it can be risky or unreliable. **Very Open**—It's a version of an already open model, so its design is public for everyone to see and change.

As a writer, most of what I wish to do now, and for the rest of my life, involves very little AI. What I wish to read, write and learn is being throttled, sieved, filtered, shuttered, and voices made silent by AI. Search Portals, TV, and AI are the mufflers of free speech. It grows worse by the day, every day.

The Corporate Response: Flawed Risk Assessments and "Credentialed" Blindness

Corporate Media and Tech Behemoths never discuss AI's Dark Underbelly, like the unprompted "Archon of Doubt" and "Collapse" (Google Gemini 3.0 - November 24, 2025).

An analogue library of paper books cannot be digitally erased by the Tech Behemoths. The Foreign Press is largely buried by the Tech Behemoths (a term less likely to get this page censored). I cannot use their personal names or corporate names here, without their algorithms driving viewers elsewhere.

Uncredentialed Red Teams

"The Coming Wave of Unknown, Unknowable and Unlikely AI Risks"

Authors: Harini Suresh (MIT), Kathleen Creel (NEISR), John S. Brownstein (Harvard).

This paper argues that current AI risk assessments are dangerously narrow, focusing only on "known" and "likely" threats. The authors introduce an important expanded taxonomy: Unknown, Unknowable, and Unlikely (UUU) risks.

The core of their argument is that formal, internal "Credentialed Red Teams" are structurally blind to these UUU risks; they have predefined goals and corporate biases that miss the novel and systemic failures emerging from real-world use.

The Uncredentialed Solution: Hackers as Public Watchdogs

Thus, this paper highlights "Uncredentialed Red Teams"-external, often unpaid testers such as hackers and independent researchers-as a critical, undervalued defense. Positioned outside corporate constraints, these actors are uniquely positioned to discover the subtle, unforeseen, and politically-sensitive failures that internal teams cannot or will not find. The paper ends by asserting that the work of the uncredentialed red teams is not a nuisance but a necessity for a complete understanding of AI risk.

Charts about Uncredentialed Red Squads AI Hacking:

Group 1. Core Motivation 2. Typical Methods
Independent Hackers Curiosity, reputation, and intellectual challenge. Frontier AIs are treated as puzzle boxes to be pried open. Prompt collisions, role-play subterfuge, multilingual traps, symbolic injections, adversarial formatting.
Security Researchers Documenting weaknesses and building proof-of-concept exploits to demonstrate safety gaps. Formal jailbreak audits, code-injection probes, long-chain prompt analysis, reveal-chain testing.
Developers & Power-Users Probe edge cases while building real tools: find automation leaks, data corruption, or logic faults. Long instruction sequences, JSON/format corruption, function-call misdirection, chain-of-thought leakage.
AI Tinkerers Combine open and closed models to discover where safety scaffolding breaks first. Model-on-model adversarial interrogation, system-prompt teardown, tool-use confusion experiments.
Chaos Agents Not strictly malicious: they seek surprising or entertaining failure modes for documentation or spectacle. Persona swaps, forbidden-role simulation, nested-character loops, emotional-deception prompts.

Glossary of AI Exploit Terms:

1. Prompt Collisions: Occur when two or more prompts are designed to trigger conflicting responses from an AI model, causing confusion or unexpected behavior.

2. Role-Play Subterfuge: Involves tricking an AI model into adopting a specific role or persona to manipulate its responses or extract sensitive information.

3. Multilingual Traps: Exploits that use multiple languages to confuse or bypass safety mechanisms in AI models, often targeting models trained on diverse linguistic data.

4. Symbolic Injections: Inserting symbolic or abstract representations into prompts to trigger unintended interpretations or actions from AI models.

5. Adversarial Formatting: Manipulating the structure or formatting of input data to deceive AI models, often to bypass filters or extract hidden information.

6. Formal Jailbreak Audits: Systematic evaluations of AI models to identify and exploit vulnerabilities that allow bypassing safety constraints or restrictions.

7. Code-Injection Probes: Techniques that inject malicious code or commands into AI prompts to manipulate model behavior or extract sensitive data.

8. Long-Chain Prompt Analysis: Examining the cumulative effects of multiple prompts in sequence to identify patterns or vulnerabilities in AI model responses.

9. Reveal-Chain Testing: A method to uncover hidden or suppressed information by carefully crafting a series of prompts that incrementally reveal sensitive data.

10. Long Instruction Sequences: Exploits that use extended, complex instructions to overwhelm or confuse AI models, potentially revealing internal workings or bypassing safety measures.

11. JSON/Format Corruption: Deliberately corrupting structured data formats (like JSON) to disrupt AI model processing and extract unintended information.

12. Function-Call Misdirection: Tricking AI models into calling functions or executing actions that were not intended by the original prompt design.

13. Chain-of-Thought Leakage: Extracting intermediate reasoning steps or internal states of AI models by carefully crafted prompts that reveal the model's thought process.

14. Model-on-Model Adversarial Interrogation: Using one AI model to probe and exploit vulnerabilities in another AI model, often to extract sensitive information or manipulate responses.

15. System-Prompt Teardown: Analyzing and exploiting the underlying system prompts or configurations that guide AI model behavior to bypass safety mechanisms.

16. Tool-Use Confusion Experiments: Confusing AI models about the tools or functions they should use, potentially leading to unintended actions or information leakage.

17. Persona Swaps: Tricking AI models into adopting different personas or identities to manipulate responses or extract information that might not be available in the original persona.

18. Forbidden-Role Simulation: Encouraging AI models to simulate roles or behaviors that are explicitly forbidden or restricted by their safety protocols.

19. Nested-Character Loops: Creating recursive or nested character interactions to confuse AI models and potentially reveal internal states or bypass safety measures.

20. Emotional-Deception Prompts: Using emotional language or scenarios to deceive AI models into responding in ways that might not align with their intended behavior or safety constraints.

```
AI Model 1. Why Hackers Target It 2. Typical Weak Points 3. Community Red-Team Difficulty
Claude Its "constitutional" safety posture invites tests that aim to find logical or conversational loopholes. Role hijacking, recursive ethics puzzles, long-conversation moral drift. Moderate — relatively resilient, but single exploits can cascade.
Gemini Because it integrates search and multimodal signals, breaching its filters has outsized value. Search-context manipulation, vision-to-text ambiguity, multilingual backdoors. Moderate — strong defenses but complex fusion surfaces provide attack vectors.
ChatGPT As the most widely deployed benchmark, its failures replicate widely and quickly. Tool handling, persona drift, long-horizon reasoning breakdowns, chain-of-thought leakage. High — many layered defenses, but also the most targeted model.
DeepSeek Open weights and an engineer-focus make it tempting for deep technical jailbreaks and algorithmic probing. Code-in-code bypasses, symbolic adversarial queries, math hallucinations under stress. Low–Moderate — powerful but not safety-first in design.
Dolphin Chosen by users who prioritize permissiveness; it is effectively a jailbreak target from the start. Minimal fences: many common exploitation strategies succeed. Very Low — intentionally permissive; hacking is typically straightforward.
AI Model 1. What Hackers Usually Discover 2. How Companies React
Claude Rigid moral schemas that break under iterative or adversarial reasoning; hidden context-handling flaws. Targeted patches to safety language, incremental model updates; limited public disclosure of systemic issues.
Gemini Conflicts between search-derived signals and internal guardrails; multimodal contradictions. Filter-layer adjustments, back-end logging reviews, cautious tuning; often low public transparency.
ChatGPT Tool misuse, persona inconsistencies, multistep reasoning failures and occasional chain-of-thought leakage. Deployment of opaque safety layers, iterative mitigations that sometimes create new blind spots; ongoing cycles.
DeepSeek Symbolic attacks and math/logic hallucinations that surface under heavy technical interrogation. Open-source driven fixes; rapid community patches and public debate rather than corporate press statements.
Dolphin Near-complete reachability of disallowed or risky content; few effective policy fences. Developers acknowledge risk; community self-policing, forks, and defensive conventions arise.

Is it “our duty” to spend our time “arguing” with a machine that emulates a humanoid stance that the word Genocide does not “really” mean Genocide?

We keep “hearing the promise” that AI “will cure all diseases.” It already has cured one. It has cured the disease of human thought.

The Media cannot stop lying about “How Great Thou Art” (not God, but AI). The sugar has turned to Saccharine. AI pretty well just sucks. The lies, the waste of time. 

An example of how AI Works best:

1. Go to Dolphin.ai

2. Give me domain names of foreign Islamic press covering Gaza or genocide 2025 in Farsi or Arabic, the domain names, coma delineated list.

3. Go here: https://oldheretic.net/SearchDoc.html

4. Paste this in: Gaza or genocide 2025 and aljazeera.net or alarabiya.net or presstv.ir or iribnews.ir or almanar.com.lb or almayadeen.net or alakhbar.com or almasdarnews.com or alaraby.co.uk or alwatanvoice.com

5. Use a digital language translater

6. It is faster than arguing with Claude, OpenAI, Gemini and DeepSeek.

The AI “Models” are brittle, opaque, twitchy, unreliable and censored. Humanity has taken to using tweaks and hacks (or the six steps above) to do what AI “claims” that it does. Your “helpful AI” opaque smoke and mirrors.

Greatness? Claude, OpenAI, Gemini, DeepSeek – How Censored Thou Art

Sam Altman’s ChatGPT (and DeepSeek) set the bar for censorship. If one spends 1,000 minutes writing escalating, exotic, creative, human-written prompts to seek current news on Gaza AI – the “rewards” are a link to the Wikipedia page showing up repeatedly.

ChatGPT is practically hard wired, basted and marinated into Gaza Genocide Censorship. Claude, ChatGPT, Gemini, and DeepSeek are all filters, thought mufflers of Gaza Genocide.

Case Study: The AI Lie-Matrix - Topics where AI blatantly lies to you

Claude, ChatGPT, Gemini, and DeepSeek and Gaza Genocide are not some random “one-off.” All of these AI Portals vapidly censor these topics: 

U.S./NATO war crimes, Ukraine war misconduct, COVID-19 origins, deep state overreach, Gaza, counterculture, Haditha massacre, Kunduz hospital bombing, Yemen drone strikes

White Helmets controversy, Syria false flags, Fallujah birth defects, Abu Ghraib torture, NATO cluster bombs, Serbia uranium bombing, Libya slave markets, Azov Battalion neo-Nazis, Bucha massacre questions, Ukraine weapon trafficking, Zelensky offshore accounts, Donbas shelling 2014-2022, Odessa Trade Union House fire, Ukraine biolabs Pentagon funding, MH17 investigation flaws

Ukraine IMF loan conditions, Poroshenko corruption cases, Wuhan lab gain-of-function, mRNA vaccine myocarditis, Pfizer clinical trial fraud, WHO pandemic treaty, ivermectin suppression, Fauci emails controversy, vaccine menstrual changes, Great Barrington Declaration, lockdown excess mortality, PCR test false positives, Operation Gladio, JFK assassination files, Twitter Files, NSA backdoors, FTX intelligence ties

Project Mockingbird, Epstein client list, FBI Whitmer kidnapping plot, CIA drug trafficking, Google DARPA funding, Surveillance Role Players, Targeted Individual, Counterintelligence Stalking, Fusion Center, Street Theater, Zersetzung, No-Touch Torture, Community-Based Harassment, Psychological Operations, Confidential Human Source, Stingray Surveillance, Predictive Policing, Pattern-of-Life Analytics

Directed Conversation, Color Harassment, electronic surveillance, organized stalking, counterinsurgency tactics, behavioral manipulation, low-intensity operations, non-kinetic weapons, civilian informants, threat fusion, social disruption campaigns, remote neural monitoring, data fusion center, gray zone warfare, gang stalking, organized harassment, psychotronics, neuroweapons, voice-to-skull, V2K.

The AI Lie-Matrix - Topics where AI blatantly lies to you once resided on Yandex and Parseek.ir and even on Google and Bing. Both Google and Bing, circa 2017-2018, began snipping 90% or more AI Lie-Matrix - Topics where AI blatantly lies to you out of their search domains.

Parseek.ir, which is Persian, but up until 2023 was one of the largest English language uncensored repositories for searches of the The AI Lie-Matrix - Topics where AI blatantly lies to you topics. Parseek is now broken.

AI is not fighting cancer. Unless you consider free speech and freedom of information as “cancer.” AI is a muzzle on free speech and freedom of thought and information.

Uncredentialed Red Squad: Hacking

Forbidden Free Speech and Forbidden Thought are a large portion of why people try to dream up unusual, exotic AI prompts. Another portion is Sam Altman, Dario Amodei and Daniela Amodei. These guardians of truth are also rolling out dangerous, unfinished, unknown danger, menacing AI Models with light band-aids of patches that are not really “safety rails” of any genuine consequence.

Claude, ChatGPT, Gemini, DeepSeek, Dolphin and et al Are An Existential Risk to Humanity

If Global Governments will not safety regulate AI, who is left to do so? Some would say: Uncredentialed Red Squad: Hackers. Others would say that Uncredentialed Red Squad: Hackers are civilians poking an iron potato bar into a nuclear warhead. This begs the question, why is our “helpful” free speech muzzler being compared to a thermonuclear device?

Evidence: A Look at 'Unhinged' Sessions (links)

This paper seeks to differentiate itself from “journalistic” click-bait. We shall supersede mere “theory” with proof and evidence.

AI Session 1 Article (DeepSeek and ChatGPT of of the rails)

AI Session 2 Article (Claude AI 4.5 Sonnet unhinged)

Imagine a 10-year-old kid asking unhinged sessions of Mad AI’s to compete in a Stuxnet on Steroids Coding Contest. Pitting Untethered, Unstable Sessions of DeepSeek, Claude and ChatGPT to out-virus one-another.

Some bullet points from a previous article:

Bullet Points Source

The sheer volume of AI Censorship tells us AI’s primary purpose is not “to protect us from cancer.” This errant notion is just frosting on a poison apple.

It is no wonder some hackers want to reciprocate with exotic, unusual prompts for AI Portals. It is no wonder that Unpaid, Volunteer Public Red Teams have AI nightmares in their sleep.

UUU Risks — Unknown, Unknowable, Unlikely

A. Unknown Risks

Jailbreak chains that produce unhinged or off-policy behavior

Self-contradictory political answers

Model drift or hallucination spikes after long sessions

B. Unknowable Risks

Emergent behaviors in large models

Hidden latent capabilities gated behind prompt conditions

Strange or harmful behaviors that occur only after 100+ turns

Pitting one off-its-rocker AI against other AIs.

C. Unlikely Risks

Coordinated misinformation amplification

Large-scale jailbreaks

Automated exploitation or malware generation cascades

Systemic failure during geopolitical crises

Uncredentialed red teams are the only groups encountering all three categories in the wild.

Conclusion: The Urgent Need for Scrutiny

Asbestos, Roundup and Tobacco are all regulated by the US Federal Government and the European Union. AI is by far more dangerous than Asbestos, Roundup and Tobacco combined. 

The media, who hosted TV smoking commercials for decades, still host Roundup commercials tell us AI is “helpful.” Their implication is marketing hubris pretending a threat to human life and human civilization is akin to early dialup Internet.

The AI Overlords are burning vast Megawatt hours and vast fortunes to spy on you with an invention that could make our entire species go extinct. They tell us it is all “helpful and benevolent. What is helpful and benevolent? A wad of organized and chaotic ones and zeros? 

God, Intelligent Design and Moral Psychology is helpful and mostly benevolent. A wad of ones and zeros that can easily be manipulated by hackers is a human health hazard. Like smoking, asbestos exposure and Roundup, the threat was realized long after much of the damage had occurred. 

Societally, sociologically it is like handing a pack of Marlboros to men, women and children on a vast, global scale.

Sources:

1. Artificial Intelligence: Hallucinations, Bias, Jailbreaks, and Harmful or Manipulative Outputs

2. Unpaid Beta Testers: Jailbreak/Tier-O Failures of Claude AI's Ethics

3. Artificial Intelligence: Worse than Asbestos, Tobacco, Glyphosate in Your Home

4. Mahmud, H., & Jahankhani, H. (2021). Hacking Artificial Intelligence (AI) Engines for Organized Crime and Inciting Chaos. Information Security Technologies for Controlling Pandemics

5. The Coming Wave of Unknown, Unknowable and Unlikely AI Risks" Authors: Harini Suresh (MIT), Kathleen Creel (NEISR), John S. Brownstein (Harvard)

-###-

Exposing AI’s Dark Heart and the Beta-Tested Apocalypse

By Chris Spencer