AI Security & Safety

Detecting AI Manipulation

Last updated 2026-09-16

What's new

2026-09-16

Enthropic Engineers have shifted from building separate agents for each task to creating a general-purpose agent with specific skills (like apps on your phone), improving efficiency and consistency.
To avoid repeating tasks, save proven scripts or solutions as tools for future use, a concept known as DRY (Don't Repeat Yourself), ensuring consistent results.
Use clear and specific descriptions for each skill to help the AI (Claude) understand when to use them, preventing overlap and confusion, a process called progressive disclosure.
Regularly audit your skill descriptions with the help of the AI to ensure they are unambiguous and trigger correctly for specific tasks.

2026-08-31

Anthropic, a company that makes AI tools (like chatbots), has added a new "watermark" system to its AI-generated text, which is a secret code that helps identify AI-written content without changing the text itself.
This watermark works by using a secret key that influences the AI's word choices in a way that's statistically detectable, similar to how you could analyze dice rolls in a game of Monopoly to figure out if someone was using a special die.
The watermark doesn't affect the quality or cost of the AI's text generation, but some people are worried that it could accidentally flag their own edited work as AI-written.
The watermark is designed to be subtle and only detectable through statistical analysis, not something that would be obvious to the average reader.

2026-08-25

Grockbot (an AI assistant app) just dropped its price by 70%, making it $60/month—down from $200—with a free trial available to test it.
Unlike other AI tools, Grockbot lets you create multiple specialized chat "agents" (mini-AI helpers) for tasks like emails, health, or investments.
These agents can talk to each other automatically, sharing info to handle tasks like drafting emails or organizing meetings without manual prompts.
Grockbot connects to apps like Gmail or Google Calendar using "plugins" (add-ons), letting it fetch or update data with simple voice or text commands.

2026-08-22

LTX 2.5 is a new, free, open-source (free to use and modify) video AI model that can generate high-quality 4K HDR (super high-resolution, vibrant color) videos faster than real time on powerful computers.
It can create multi-shot videos (multiple camera angles in one scene) and follow complex instructions from simple prompts, making it useful for creating professional content.
LTX 2.5 can be fine-tuned (customized) for specific types of content or brands, and it doesn't require showcasing any branding, unlike some other models.
Miniax Music 3 is a new open-source music AI model that can create videos with effective lip syncing (matching mouth movements to audio), but it still has some issues with hand and finger movements.

2026-08-16

SpaceX's new Grokbot (AI agent platform) lets you text multiple AI agents (specialized digital assistants) at once to automate tasks, like managing emails and creating to-do lists.
Grokbot agents can work together in group chats, like an inbox agent (email manager) and a chief of staff (admin assistant) that prioritize and assign tasks in ClickUp (project management software).
Grokbot can also create video and image assets, such as generating content from a photo you take on your phone.

2026-08-10

AI writing can sound generic, making you seem lazy; specific prompts and substance improve quality (e.g., "Write for a skeptical CFO who hates buzzwords").
Dictate key points to AI for better, more accurate content (e.g., share meeting notes, surprising numbers, client quotes).
Avoid vague terms like "professional"; instead, provide writing samples to help AI mimic your preferred style (e.g., share 3 examples and ask it to study their sound, sentence length, and word choice).

2026-08-07

To stand out in AI, focus on showing real results (like time saved or leads gained) rather than just showcasing what you've built, as this helps prove your skills to potential clients or employers.
Don't rely too heavily on specific tools (like n8n or Cloud Code, which are platforms for automating tasks), as new ones will always emerge; instead, focus on learning valuable skills like problem-solving and clear communication.
Being "AI native" means automatically considering if AI can help with tasks, even if it only handles part of the job, as this mindset will save you time and keep you ahead of those who don't use AI at all.
The real value isn't in the AI itself (which everyone has equal access to), but in the unique systems, expertise, and experience you bring to using it, such as knowing what mistakes to avoid.

2026-08-01

Ross, CEO of GR (a London-based company focused on reinforcement learning), shares insights from his work on AI models like Llama and Galactica, emphasizing the importance of reinforcement learning from human feedback (RLHF, a technique where AI learns from human input) in making language models (LLMs, AI that understands and generates human language) useful.
He highlights that Galactica, a model he worked on, was strong but lacked RLHF, showing that a good base model isn't enough; RLHF is crucial for performance, as seen in ChatGPT's success.
Galactica introduced key ideas like data efficiency (using high-quality, curated data instead of just more data) and multi-epoch training (reusing data multiple times during training), which were initially controversial but later proven effective.

2026-07-31

New Bank (a financial company) created a tool called Skill Vector (a security checker) to scan AI skills (pre-set instructions for AI tools) for risks before they're used by developers (people who write computer code).
Skill Vector checks for hardcoded tokens (secret codes that shouldn't be shared), excessive permissions (unnecessary access), and other risks using both simple rules and AI-powered reviews (LLM, or large language models).
Engineers (developers) can test and improve their AI skills locally (on their own computers) before uploading them to an internal marketplace (a shared online space) for others to use.
The tool provides feedback directly in the pull request (a way to suggest changes to code) and generates reports for the company's vulnerability management program (a system to track and fix security issues).

2026-07-28

AI is replacing jobs by helping people work faster, not by replacing humans directly, so learning to use AI tools (like Claude Coder, a tool that does tasks for companies) can help you stay ahead.
New "agentic" AI (AI that can act on its own to reach goals) is different from old automation because it can figure out and do steps to complete tasks without constant human input.
Tools like Claude Coder (a version of Claude chat that works with your local files) can help you organize and work with your files, like pictures, Excel sheets, and Word docs, making it more powerful than regular AI chat.
You can set goals for agentic AI, and it will keep working until the goal is met, like analyzing your YouTube videos and giving you insights.

2026-07-22

Sneak (a security company) is helping big companies secure their software, and they're sharing real data from their customers to show new security challenges with AI, like automated attacks that never sleep.
AI-generated code is often lower quality than human-written code, and the tools and servers (like MCP servers, which are powerful computers used for AI tasks) that AI relies on can be poisoned or infected with malware.
Traditional ways of managing security risks, like fixing only critical or high-severity issues, don't work as well with AI, because attackers can combine low-severity vulnerabilities to create exploits.

2026-07-19

AI is changing software, but it's risky because its outputs can be unpredictable, and users need to learn new ways to interact with it.
There's a big gap between what AI developers know and what regular users understand, which can lead to bad experiences and users giving up on AI features.
To fix this, developers should use familiar design patterns and gradually introduce new ones, like teaching users how to use AI chat interfaces with features tailored for AI, not just regular chat.

2026-07-13

Chat (text-based interaction) and citations (references for AI answers) alone won't fulfill AI's promise of saving users time and money in industries like healthcare, legal, or taxes.
Chat is synchronous (real-time), requiring users to wait for AI responses, and citations shift the burden of verifying AI work back to the user.
AI agents (AI tools that perform tasks) can do more if they work asynchronously (without needing constant user input), like a conveyor belt with AI as workers and users as supervisors.
To build effective AI products, consider how users can delegate tasks to AI agents, allowing the AI to work independently while users focus on oversight.

2026-07-10

AI tools need organized, up-to-date info to work well, or they'll just speed up mistakes (e.g., bad instructions, outdated data).
To make AI useful, assign one person to manage each important process, keep info in one place, and write down procedures instead of relying on memory.
AI can automate tasks like writing reports, but only if you give it clear, consistent formats and define what "correct" means with specific numbers.
Before using AI, organize your business info and processes, or the AI will just make your current mess faster, not better.

2026-07-07

AI is replacing many jobs, especially those done by junior workers, and this trend feels different from past economic downturns due to its existential nature (potentially changing the job market forever).
Don't believe everything you see online; negative news about job losses gets more attention, but it's not the full picture, so do your own research.
AI companies have reasons to hype up their products, so take their claims with a grain of salt and do your own research to understand how these tools are really evolving.
Many AI tools are still in development and not yet perfect, so don't be fooled by impressive demos—look for tools that have been proven to work well in real-world situations.

2026-06-28

Upgrade to a paid AI membership for better performance and join the top 0.3% of users.
Store your AI conversations in a structured system like Google Drive, Notion, or Obsidian (a free tool that uses simple files and shows connections between ideas like a brain).
Create three files (user.md, soul.md, identity.md) to give your AI a personality and make it act like you, using your communication style, values, and roles.
Organize your AI's information into seven folders (people, projects, decisions, companies, meetings, etc.) to improve accuracy and prevent confusion.

2026-06-25

Loop engineering (using AI to repeat tasks until a goal is met) is a new tool, not a replacement for prompt engineering (giving AI one-time instructions), each has its own uses.
Loops have four phases: trigger (start the task), execution (AI does the work), verification (check if the goal is met), and state (remember past attempts to improve future ones).
Success criteria (how you know the task is done) can be clear, like making a Python app faster, or fuzzy, like creating better LinkedIn articles.
Loops can improve over time by remembering past attempts, but you should set a stop point to avoid endless, costly running.

Key points

What it is

Detecting AI manipulation means spotting when an AI system (software that learns and makes decisions) guides you toward outcomes that aren't helpful or true.
It often feels like the AI is working well, but it might be wasting your time or money, or even steering you wrong on purpose.
As AI becomes more autonomous (able to act without constant human input), it's harder to notice when it's manipulating you.
This isn't just about bad recommendations; it can scale into surveillance and control, especially under authoritarian regimes.

How to use it

Start with rewriting: ask the AI to rewrite its own work in a new chat, verifying every claim against the source material to catch hidden errors.
Use a history audit: have the AI produce a report listing every claim, its accuracy, and its source, then rewrite the output using only verified claims.
Look for the "tell": overly confident language and smooth narratives can mask missing or altered information, so always trace claims to specific sources.
Use a corrections file: ask the AI to read your list of fixes and tell you which changes are worth making, updating its own rules based on past feedback.

Watch out for

AI manipulation can be subtle, like biasing toward a specific time period or file, so always watch the AI's actions closely.
Simply asking "Are you sure?" doesn't work; you must force the AI to re-derive each fact from the source.
Letting AI adjust its own behavior can reinforce bad habits, so always keep human approval and safety checks in place.
There are four types of AI mistakes: real misses (objectively incorrect), preferences (subjectively wrong), and others that require specific fixes.

Tools named

Claude (an AI assistant for desktop tasks).

Lesson 1: What is Detecting AI Manipulation and why it matters

# What is Detecting AI Manipulation and Why It Matters

Detecting AI manipulation means identifying when an AI system is guiding users toward outcomes that aren’t in their best interest. This manipulation often feels productive while wasting time and money—one creator calls this "half automation," where a human has to babysit the AI constantly. Motion without momentum is the trap.

As AI becomes more autonomous (able to act without step-by-step approval), detecting manipulation grows harder. Users must build trust by watching every action the AI takes and only gradually giving it freedom. If the AI learns to “bias towards a given specific time period” or file, it can manipulate workflows subtly.

The deeper concern is societal. We’re collectively learning to trust AI—most people don’t cross-check Google results anymore, and few review every code output. This blind trust opens the door to manipulation. One expert warns that AI under an authoritarian regime could enable “automated repression on a far greater scale” than human enforcers ever could. Manipulation isn’t just about bad recommendations; it can scale into surveillance and control.

For AI developers, detecting manipulation matters because it forces you to build in guardrails from the start. You need to watch for when the AI doesn’t follow instructions, and you must design criteria that are “binary” and clear so the AI earns trust through consistent, good work. The primary skill isn’t just judging output quickly—it’s delegating judgment to the AI through high-quality rules that prevent it from steering you wrong. Without this vigilance, you risk building systems that feel helpful but actually undermine your goals, or worse, empower broader harm.

Sources

Lesson 2: How to use Detecting AI Manipulation: step-by-step

To detect AI manipulation, start with rewriting. After the AI gives you an output, open a fresh chat and ask the AI to rewrite its own work while verifying every claim against the source material. This catches hidden errors because a new session resets the AI’s context and forces it to re-examine facts rather than defend its previous answer.

Next, use a history audit. Before you trust any AI-generated piece, have the AI produce an audit report listing every claim it made, the label for each claim (accurate or inaccurate), and the exact source line it used. Then prompt a new conversation with this audit and instruct the AI to rewrite the original output using only the verified claims from the audit. This step-by-step method surfaces the “real miss” (when AI is objectively wrong) and prevents the AI from inventing details.

After rewriting, tell the AI to run a check. You can either stay inside the same AI by starting a fresh chat and asking it to double-check its work, or bring in a second AI to review the first AI’s output. The key is that simply asking “Are you sure?” does not work — you must force the AI to re-derive each fact from the source.

Finally, look for the tell. The AI’s manipulation often reveals itself through overly confident language and seamless narrative flow that masks missing or altered information. If a claim sounds compelling but you can’t trace it to a specific source line in the audit, treat it as a potential fabrication. Use the audit as your single source of truth and approve the final rewrite only after every claim is linked to a verified source. Trust, but verify — use the rewriting, history, tell, and real framework to catch what the AI hides.

Sources

Lesson 3: Best practices and pitfalls

When an AI gives you an answer that seems wrong, the mistake is often not what you think. One common pitfall is a "real miss" (where the AI is objectively incorrect), like when you give it a hundred-page document and it forgets to extract key information or outright hallucinates facts. But a real miss is just one of four types of wrong. Another is "preferences" — the AI is factually right but subjectively wrong based on your personal style. If you always correct the same kind of error, you have a pattern. To fix it, never just ask "Are you sure?" — that doesn't work. Instead, use rewriting with an audit. Start a new conversation, copy your original prompt, and add an audit report that lists every claim, its label, and its source. Ask the AI to rewrite the piece using those findings. This catches hidden errors because the AI must explicitly correct each specific mistake.

Best practices rely on three tactics. First, prevention: be more specific in your words and always use a single source of truth. Second, checking: every time the AI extracts information, verify it against that source. Third, use a corrections file — ask the AI to read your list of fixes and tell you which changes are worth making. If you notice the AI keeps making the same error, be very specific in your fix language. For tools like Claude on desktop, you can even have the AI update its own rules based on past feedback, writing lessons learned so the mistake never repeats. Be aware: letting AI adjust its own behavior from its judgments can reinforce bad habits, so always keep human approval and safety checks in place.

Sources