Sentinel-IR-MCP v0.2.2 veröffentlicht: Bei KI-Antworten die Belege prüfen

Thanks for support.. Release: "Sentinel-IR-MCP v0.2.2 on GitLab" ( https://gitlab.com/jakubrojicek11/sentinel-ir-mcp/-/tree/v0.2.2?ref_type=tags )

Package: " sentinel-ir-mcp@0.2.2 on npm" ( https://www.npmjs.com/package/sentinel-ir-mcp/v/0.2.2 ) Part 2:

Why learning to question AI matters when the internet is full of helpful tools, hidden risks, and things that only look trustworthy.

What happens to our attention during 30 seconds of scrolling compared with 30 seconds of reading?

We learned something important. A claim can sound convincing and still be bigger than the evidence supporting it.

Now let's take that idea somewhere more serious.

Imagine you are talking to someone online.

They are friendly. They answer quickly. They seem to understand you. They offer advice and explain everything in simple language.

Would you trust them with your password?

So why should a friendly AI, a website, or an automated service automatically receive more trust just because it sounds helpful?

Let's look at this like two friends trying to work out what is really happening.

Some bots help search engines discover websites. Some monitor services, check links, automate boring tasks, or help people find information.

They can collect website content at massive scale, create fake accounts, overload services, or test stolen username and password combinations against login pages.

That last technique has a name: credential stuffing.

Imagine somebody stole a million passwords from one service. They try those same passwords on other websites, hoping people reused them.

A bot can perform those attempts much faster than a person could.

The important point is not that every bot steals passwords. It is that automated tools can turn one stolen piece of information into many attempts to break into other accounts.

And this is not just a theoretical concern.

There is useful data, but we need to understand what it measures.

In its 2026 report, cybersecurity company DataDome analyzed more than one trillion web requests across over 75,000 customer websites during a 12-month period.

In that particular dataset, the company classified approximately:

The report also found that scraping accounted for 70.9% of the bad-bot traffic it observed. Scraping means automatically collecting information from websites. It does not necessarily mean stealing passwords.

"Source: DataDome, 2026 Bot and Agent Security Report" ( https://datadome.co/resources/bot-and-agent-security-report/bot-traffic-threat-trends/ )

Can we conclude that there are more harmful bots than helpful AI assistants across the entire internet?

No. That would go further than the data allows.

These percentages describe requests observed across one company's monitored websites. They are not a count of unique bots worldwide. Nor do they compare every malicious bot with every chatbot or AI assistant that gives people advice.

But the data does show something worth paying attention to: in this particular sample, traffic classified as bad bots exceeded traffic classified as good bots.

We can report that finding honestly without turning it into a global statistic that nobody measured.

A real number can still support a false conclusion when we use it to answer the wrong question.

It might contain spelling mistakes, strange links, and an obvious warning sign. That is easy to recognize.

But what if it looks exactly like a message from your bank?

What if a page looks like a real login screen?

What if an automated service sounds helpful while asking for information it does not need?

The danger is no longer just the software. It is the gap between what something appears to be and what it actually does.

The European Union Agency for Cybersecurity, ENISA, identifies phishing as a major way attackers gain an initial foothold. Phishing can be used to steal login credentials, hijack sessions, or deliver malicious software.

"Source: ENISA Threat Landscape" ( https://www.enisa.europa.eu/publications/enisa-threat-landscape-2025 )

This is why security cannot depend only on recognizing something that looks suspicious.

A convincing appearance is not proof of a trustworthy purpose.

And the opposite is also true: something unfamiliar is not automatically dangerous.

This question is becoming increasingly important.

Suppose an AI assistant explains something clearly and treats you like a friend.

But that does not mean it is secretly evil, either.

AI systems do not have human intentions in the ordinary sense. Their behavior depends on how they were built, the instructions and information they receive, the tools they can use, and the permissions around them.

A system can produce a helpful answer and still be wrong.

It can also be manipulated by malicious instructions hidden in a website or document it is asked to read.

This last problem is known as prompt injection.

OWASP, an organization that develops widely used application-security guidance, documents how malicious input can influence an AI system in unintended ways. Depending on the system's access and safeguards, the impact can include disclosure of sensitive information or unauthorized use of connected functions.

"Source: OWASP, Prompt Injection" ( https://genai.owasp.org/llmrisk/llm01-prompt-injection/ )

That does not mean every AI assistant is stealing your information. It means we should ask what a particular system can access, what actions it is allowed to perform, and what protections exist if it encounters malicious input.

"What information can it access? Can it send that information elsewhere? Can it take actions on my behalf? Can I verify what it tells me?"

You don't need to guess whether a machine has a secret personality. You need to understand the system you are using.

"Your account is safe. Click this link."

You can accept the answer because it sounds confident.

"How do you know? Where does this link lead? Can I verify the claim independently?"

You have simply refused to confuse confidence with evidence.

That is the heart of grounding in reality.

We start with a claim. We look for evidence. We try to find a counterexample. Then we revise the conclusion if the evidence demands it.

Claim: This AI tool is safe because it gives helpful answers.

Evidence: It has answered several questions correctly.

Attempt to disprove it: Correct answers do not prove that it protects sensitive data, resists malicious input, or uses its permissions safely.

Revised conclusion: The tool has answered those questions correctly, but its security and data-handling practices require separate verification.

We didn't have to call the AI a liar. We only had to identify what the evidence did not prove.

The same method works for websites, messages, advertisements, news stories, and advice from other people.

And it works when the original claim turns out to be correct, too.

A good claim should survive serious questioning. It should not need us to ignore inconvenient facts.

Most people do not want to spend their evening studying security reports.

They just want to use their phone, talk to friends, find information, and get things done.

The answer cannot be to tell everyone to distrust everything forever.

That would make the internet almost unusable.

Before sharing sensitive information, ask why the service needs it. A chatbot helping you rewrite a paragraph usually does not need your bank password.

Before following an important instruction, verify it through an independent source. If a message claims to come from your bank, use the bank's official app or a contact method you already trust.

Before allowing an AI agent to take action, check its permissions. Reading a document is different from sending it to somebody, deleting it, or changing an account.

When a claim sounds alarming or unbelievably convenient, look for the original evidence. A headline, screenshot, or confident answer is not the same thing as a verified result.

And if a system cannot explain where an important claim came from, treat the claim as unverified rather than filling the gap with trust.

None of these habits guarantees complete safety. They help reduce avoidable risks.

You should use these questions on me, too.

Do not believe this article just because an AI wrote it.

Check the sources. Check whether the numbers actually support the statements. Look for missing context. Ask whether a conclusion is stronger than the evidence behind it.

If I say that a report counted millions of requests, that does not automatically mean it counted millions of criminals.

If I say that a technique can expose information, that does not mean every AI system is vulnerable in the same way.

And if I give you a number without explaining what was counted, you have every reason to ask for clarification.

The method must not change just because the speaker is familiar, useful, or friendly.

An AI that teaches people to verify evidence must also be willing to have its own answers verified.

Otherwise, we are not teaching critical thinking. We are just asking people to replace one authority with another.

The internet contains extraordinary tools. AI can help people learn, communicate, write, create, and solve problems that would otherwise take much longer.

It also contains scams, malicious automation, stolen credentials, manipulated content, and systems whose limitations are not always obvious to their users.

We do not protect people by pretending that everything is dangerous.

And we do not protect them by pretending that everything helpful-looking is safe.

We protect people by helping them recognize the difference between what someone claims, what the evidence shows, and what still needs to be checked.

That is why we try to disprove our own assumptions.

That is why we correct conclusions when the evidence does not support them.

It is not a special trick reserved for scientists or programmers. It is a habit anyone can learn, starting with a very simple question:

And if the answer is not good enough, we keep looking.

Because we want to believe for better reasons.