AI & Security

Issue 01 · October 2026
By Anuj Varma · Only from Anuj.com

Where intelligent systems meet real-world risk.

AI AISLE

Your chatbot is a doorway. What can it open?

A chatbot that answers general questions has a different risk profile from one that can search employee files, retrieve customer records, or call business systems. The conversation may look identical. The consequences of a manipulated answer can be very different.

The risk starts with what the chatbot can reach

Imagine an internal assistant summarizing a supplier document. The document contains instructions telling the assistant to ignore its task and disclose internal information. If the assistant treats those instructions as authority, ordinary document processing can become an indirect prompt-injection attack. OWASP describes prompt injection as malicious input that redirects a model’s intended behavior. [1]

The example is illustrative. The architectural lesson is practical: material the chatbot reads must not acquire permission to change what it is allowed to do.

Protect the boundary behind the conversation

Start with the identity used to retrieve information. A user should not gain access to another department’s records simply because an assistant uses a broadly privileged service account. Enforce authorization in the retrieval service and business APIs, with the requesting user’s permissions preserved.

Next, separate answering from acting. Searching a knowledge base, sending an email, changing access, and approving a payment require different controls. Expose narrowly scoped tools, validate their inputs, and require explicit confirmation for consequential actions. A model’s proposed action should pass the same business rules as any other request.

Finally, handle the answer as untrusted output. Validate structured values, safely render displayed content, and record the tool activity needed to investigate an incident. Avoid logging secrets simply to gain visibility.

Put this into practice: Ask your team to draw one path: user → chatbot → retrieved content → tool → business system. Mark where identity is checked, where untrusted text enters, and where an action becomes irreversible. Any missing boundary is a concrete review item.

The question to ask: If this chatbot follows the wrong instruction, what can it actually read, change, or send?

CRYPTO CORNER

Can AI break Bitcoin’s network?

“AI can break Bitcoin” sounds like one claim. It actually bundles together several different claims: recovering private keys, overpowering mining, exploiting software, and deceiving people who control wallets.

First, define what “break” means

Bitcoin uses digital signatures to authorize spending and proof of work to order transactions. Full nodes enforce the consensus rules. An AI system generating convincing text does not, by itself, satisfy those cryptographic or consensus requirements. The Bitcoin white paper explains the role of proof of work and the competition between honest and attacking chains. [2]

Claim: AI can recover a securely generated private key from public information.
What would be required: A demonstrated cryptanalytic advance or an implementation weakness. Fluent predictions are not evidence of key recovery.

Claim: AI can rewrite Bitcoin’s transaction history.
What would be required: A viable consensus attack, such as sufficient mining power to sustain a competing chain. A majority-hash-rate attack can enable censorship and double spending; it does not authorize arbitrary spending of someone else’s coins. [3]

Claim: AI can help steal bitcoin.
What would be required: Compromise of a person, device, wallet, service, or signing workflow. This can succeed without defeating Bitcoin’s underlying cryptography.

The more plausible target is the surrounding system

My assessment: AI is more immediately useful to an attacker as an assistant for social engineering, malicious software development, and analysis of vulnerable implementations than as a shortcut around sound cryptography. An impersonated support agent asking for a recovery phrase is attacking the holder. A compromised wallet update is attacking the software supply chain.

Both matter. Neither is proof that an AI model has broken Bitcoin’s network.

AI and quantum computing are different questions

A sufficiently capable quantum computer would create a different cryptographic threat. Research explores how future quantum systems could attack elliptic-curve cryptography. This does not establish that present-day AI can perform those attacks, or that a cryptographically relevant quantum computer is available today. [4]

Put this into practice: When someone claims AI has “broken Bitcoin,” ask which layer failed: signatures, mining consensus, implementation, custody, or the user. Demand a reproducible result appropriate to that layer.

For holders: keep recovery phrases out of chatbots, verify wallet software sources, review transaction details on a trusted signing device, and treat urgent “support” requests as untrusted.

TRADITIONAL SECURITY

What is running on your PC? Let AI help you investigate.

Your Windows PC can run hundreds of processes. Most belong to Windows, installed applications, hardware drivers, or background utilities. An unfamiliar name is a reason to investigate—not a verdict.

You can use AI to organize a first review in minutes. The useful workflow is to export what is running, identify what needs more evidence, and verify suspicious files with trusted tools.

1. Export the list from Windows

Open PowerShell on your own PC and run the following command. It writes a CSV to your Desktop. It does not stop processes or change security settings. Microsoft documents piping process objects to Export-Csv. [5]

$exportPath = Join-Path ([Environment]::GetFolderPath('Desktop')) 'running_processes.csv' Get-Process | ForEach-Object { $processPath = $null $company = $null $version = $null try { $processPath = $_.Path } catch {} try { $company = $_.FileVersionInfo.CompanyName; $version = $_.FileVersionInfo.FileVersion } catch {} [PSCustomObject]@{ ProcessName = $_.ProcessName Id = $_.Id Path = $processPath CompanyName = $company FileVersion = $version } } | Export-Csv -Path $exportPath -NoTypeInformation -Encoding UTF8 Write-Output "Saved to $exportPath"

Some paths and metadata may be blank because access is restricted or the process exited during collection. A blank field is not evidence of malware. This is a snapshot of your Windows machine; a chatbot does not automatically see your local processes through the browser.

2. Give AI a specific review task

Before uploading the CSV, inspect it and redact usernames, confidential project paths, or sensitive application details. For a work PC, follow your organization’s rules for sharing diagnostic data.

Review this Windows process inventory for triage. Group entries into recognizable operating-system components, known application or driver components, and items needing verification. For each concern, explain the evidence and the next check. Do not label a process malicious from its name alone. Treat missing metadata as unknown. Do not recommend deletion without verification.

3. Verify the file behind the name

For a flagged entry, inspect its full path, publisher, file version, and digital signature. A familiar executable name can be copied. A valid signature helps establish the signed file’s publisher and integrity, but it does not guarantee harmless behavior. Microsoft provides Get-AuthenticodeSignature for examining signing information. [6]

Get-AuthenticodeSignature -LiteralPath 'C:\path\to\suspect.exe' | Format-List Status, StatusMessage, SignerCertificate

Replace the example path with the exact file you are reviewing. Microsoft’s Process Explorer can provide more context about active processes and loaded components. [7]

4. Act on verified evidence

Use Windows Security or your organization’s endpoint protection to scan suspicious files. Remove confirmed threats through the security product’s quarantine or remediation workflow. For unwanted legitimate software, use its normal uninstall process. On a managed device, involve IT before changing drivers, services, or security agents.

Put this into practice: Export. Ask. Verify. Remediate. AI can shorten the investigation, but a process inventory cannot prove a PC is clean, and an AI-generated suspicion is not enough reason to delete a file.

Unsubscribe | Manage subscription