Gurov Digital

Research

Notes on AI, AI agents and information integrity, written by researcher Pavel Gurov

Marbled blue abstract background, liquid marble pattern — Magnific · Licensed under the Magnific Free License

8 articles

  1. Polycrisis: how the European Commission named our era, and what future historians will call it instead

    Nobody names their own era. The label a period ends up with is assigned by its outcome, and the man who popularised polycrisis has already started backing away from it

    Information integrityCounter-disinformationCognition

  2. Jacobian space: the hidden reasoning layer found inside Claude, and language prediction in the unconscious brain

    A privileged internal workspace inside a language model, and language prediction in a brain that is not conscious. Read together, they narrow the gap from both ends

    ClaudeCognition

  3. Three claims about AI that survive scrutiny: the intelligence definition, the pronoun problem, and Google's shelved LaMDA

    The term itself does not survive an academic audit, the hardest word in English for a language model is two letters long, and the company that got there first archived it

    CognitionGemini

  4. AI agent security: five layers of isolation for a self-hosted autonomous agent

    An autonomous agent with shell access and an internet connection is not a chatbot. Here is the defence-in-depth architecture I ended up with on a 2016 MacBook, and where each layer stops

    AI securityAI AgentsAI Infrastructure

  5. Prompt injection has no fix: the architectural limit of AI security

    Every AI company says it is working on prompt injection. None of them can fix it. The reason is architectural, and it is old enough to have a precedent

    AI securityAI AgentsClaudeChatGPTGemini

  6. AI security: how a Word document made Anthropic's Claude exfiltrate confidential files

    Everyone worried about installing an autonomous agent. Meanwhile the assistant most white-collar workers already trust had a vulnerability that exfiltrated confidential files in silence

    AI securityAI AgentsClaude

  7. Machine-to-machine language is emerging in latent space. Ted Chiang got the medium wrong.

    Heptapod B encoded a whole concept in one circular glyph with no sequential grammar. Three independent research efforts have now shown that AI agents communicating in raw latent vectors beat text by wide margins

    AI AgentsAI InfrastructureCognition

  8. How AI-enriched credential stuffing manufactured a mass Instagram breach that never happened

    Seventeen million records, a dark-web listing, and a wave of coverage. The database was real. The breach was not. The interesting part is how the database was manufactured

    Information integrityAI securityInstagram