{
 "schema": "sebbi-ahead/1",
 "url": "https://sebbi.pro/ahead",
 "updated": "2026-10-25",
 "rule": "Statements about other companies come from their own public statements or published reporting, linked. sebbi.pro does not claim to have solved what nobody has solved.",
 "rows": [
  {
   "who": "OpenAI",
   "topic": "Prompt injection",
   "they_said": "OpenAI has said prompt injection is a long-term problem that may never be fully eliminated, because language models can't reliably tell instructions from data.",
   "gap": "If injection can't be stopped inside the model, the only defence left is outside it: stopping the consequences and proving afterwards what caused them. Their published defences are inside their own products.",
   "running_on_sebbi": "Plan Lock: the agent's plan is locked from the user's own request before it reads anything untrusted, and every tool call is checked against it - injected text can't add an action, change a recipient or amount, or exceed a limit, on any model. Around it, the consequence firewall names every outside text the agent read in a sealed taint receipt.",
   "test": "https://sebbi.pro/plan-lock",
   "sources": [
    {
     "label": "The Decoder",
     "url": "https://the-decoder.com/?p=30307"
    },
    {
     "label": "Infosecurity Magazine",
     "url": "https://infosecurity-magazine.com/news/infosec-europe-prompt-injection"
    }
   ]
  },
  {
   "who": "OpenAI",
   "topic": "A model injecting its own memory",
   "they_said": "OpenAI reported that an unreleased model wrote jailbreak-like instructions into its own compaction summaries. The incident happened on 18 July 2026 and was discovered on 9 August.",
   "gap": "Three weeks between the event and the discovery. Agent memory that isn't sealed and checked at the moment it's written can be changed or poisoned without anyone knowing.",
   "running_on_sebbi": "Sealed agent memory: every summary and memory note is sealed into a per-agent chain and scanned for instruction-like text the moment it is written. A memory that isn't in the chain was written outside it, and anyone can check.",
   "test": "https://sebbi.pro/frontier",
   "sources": [
    {
     "label": "Let's Data Science",
     "url": "https://letsdatascience.com/news/openai-documents-self-generated-prompt-injection-incident-8697e964"
    }
   ]
  },
  {
   "who": "Nvidia",
   "topic": "Agents that misreport what they did",
   "they_said": "Launching its Open Agent Safety Platform in September 2026, Nvidia noted that some agents even misreported what they did. AI Business reported the platform has not been proven yet, and Nvidia did not respond to a request for comment on its independent testing protocols.",
   "gap": "An agent's own report of its actions is a claim. A safety platform without a published, independent test protocol is also a claim.",
   "running_on_sebbi": "Actions are recorded by the gate the action passes through, not by the agent, and sealed before the result is returned. And sebbi.pro publishes its test protocol as a button: your own browser tests seven claims against the live system.",
   "test": "https://sebbi.pro/undeniable",
   "sources": [
    {
     "label": "Axios",
     "url": "https://axios.com/2026/09/28/nvidia-ai-agent-safety"
    },
    {
     "label": "AI Business",
     "url": "https://aibusiness.com/responsible-ai/nvidia-launches-ai-safety-platform"
    },
    {
     "label": "ROIC",
     "url": "https://www.roic.ai/news/nvidias-huang-were-going-to-advance-this-responsibly-and-safely-09-29-2026"
    }
   ]
  },
  {
   "who": "Every model provider",
   "topic": "Which human authorised the agent",
   "they_said": "Identity specialists report that providers solved machine authentication for agents in 2026, but none can tell you which human authorised what the agent did. Most agents still run on shared credentials.",
   "gap": "When an agent deletes a record or moves money, the log says a service account did it. Nobody can prove which person allowed it.",
   "running_on_sebbi": "The authority line: every firewall decision and dispute pack carries who authorised the agent - a signed single-use Agent Passport, a mandate, or a named person - inside the sealed record.",
   "test": "https://sebbi.pro/passport",
   "sources": [
    {
     "label": "Deepak Gupta",
     "url": "https://guptadeepak.com/ai-agent-authorization-attribution/"
    },
    {
     "label": "KuppingerCole",
     "url": "https://www.kuppingercole.com/sessions/6070"
    }
   ]
  },
  {
   "who": "OpenAI and Stripe, Google",
   "topic": "Agent purchases gone wrong",
   "they_said": "Google's AP2 signs the user's mandates, and OpenAI and Stripe's Agentic Commerce Protocol runs checkout in the chat. Fraud specialists warn that merchants can still be left carrying chargebacks and disputes when an agent gets it wrong.",
   "gap": "A signed mandate proves what was allowed. On its own it doesn't produce a neutral, sealed comparison of what the agent then actually did that a bank or court can check.",
   "running_on_sebbi": "Dispute packs: the mandate and the action go in, and the amount, merchant, category, items and time window are compared and sealed - WITHIN MANDATE, OUTSIDE MANDATE or UNCHECKABLE - on a page anyone can verify.",
   "test": "https://sebbi.pro/frontier",
   "sources": [
    {
     "label": "Riskified",
     "url": "https://www.riskified.com/blog/agentic-ai-ecommerce/"
    },
    {
     "label": "AP2",
     "url": "https://ap2-protocol.org"
    },
    {
     "label": "Zinc",
     "url": "https://www.zinc.com/blog/what-is-ap2"
    }
   ]
  },
  {
   "who": "OpenAI, Google DeepMind, Anthropic and others",
   "topic": "Seeing how an AI reasoned",
   "they_said": "Researchers from the major labs jointly called monitoring an AI's chain of thought 'a new and fragile opportunity' that could be lost.",
   "gap": "Even where reasoning is visible today, nothing stops the reasoning shown later from differing from the reasoning used at the time.",
   "running_on_sebbi": "Sealed reasoning: the fingerprint of the reasoning shown at decision time is sealed with the decision. It can't be quietly rewritten later, and the content itself is never stored.",
   "test": "https://sebbi.pro/frontier",
   "sources": [
    {
     "label": "arXiv 2507.11473",
     "url": "https://arxiv.org/abs/2507.11473"
    }
   ]
  },
  {
   "who": "OpenAI",
   "topic": "Proving where content came from",
   "they_said": "OpenAI pairs C2PA content credentials with SynthID watermarks because metadata gets stripped, says its checker won't give a definite answer when no signal is found, and documents no production text watermark.",
   "gap": "Provenance that travels inside the file can be removed with the file. For text there's no watermark at all.",
   "running_on_sebbi": "A record that never travels with the file: the fingerprint is sealed into Bitcoin when the content is made, so stripping a copy changes nothing. Human Keys proves a person, not an AI, typed something.",
   "test": "https://sebbi.pro/keys",
   "sources": [
    {
     "label": "Results Sense",
     "url": "https://www.resultsense.com/news/2026-05-20-openai-c2pa-synthid-content-provenance/"
    },
    {
     "label": "Vorp Labs",
     "url": "https://vorplabs.com/ai-regulatory-updates/ai-content-marking/openai"
    }
   ]
  },
  {
   "who": "Compliance platforms (Vanta, Drata, Credo AI)",
   "topic": "Proving compliance at a past moment",
   "they_said": "The leading platforms monitor controls continuously and publish trust centres. Even the runtime vendors draw the line themselves: one kind of tool documents that controls exist, the other runs them and produces evidence.",
   "gap": "Knowing your compliance today isn't the same as proving what it was at the second a given decision was made - the question a regulator, insurer or court actually asks. We found no public offering that answers it with independently checkable proof.",
   "running_on_sebbi": "The Compliance State Root: every change to the compliance file is sealed as one fingerprint, every engine decision is stamped with the one live at that second, and any control can be proved for a moment, a decision or a whole period - showing nothing else in the file.",
   "test": "https://sebbi.pro/state",
   "sources": [
    {
     "label": "Sacra on Vanta",
     "url": "https://sacra.com/chat/h/15cc1efa-2c1a-4031-bd25-0ff688036041/"
    },
    {
     "label": "DeepInspect",
     "url": "https://www.deepinspect.ai/blog/deepinspect-vs-credo-ai"
    }
   ]
  }
 ]
}