BALASORE, INDIA — Working independently from a coastal town in eastern India, reportedly without venture capital, a technical staff or a patent attorney, an independent researcher has assembled a massive paper trail aimed at the core mechanics of how autonomous artificial intelligence agents act in the real world. Over the past year, Sangam Kumar Das, a self-described electronics and telecommunications engineering graduate from Utkal University based in Balasore, Odisha, reports filing 35 Indian provisional patent applications and more than 14 international applications under the Patent Cooperation Treaty, at least 10 of which he says have already been published. Whether any of those pending claims ever become enforceable patents, a costly, multi-year process across national patent offices indeed – their public release alone has now introduced a sprawling body of dated technical disclosures into databases available to technology companies and patent examiners.

Mr. Das, according to his own web page.
At the center of the portfolio is a 4 June 2026 international filing published by the World Intellectual Property Organization on 16 July 2026 under the identifier WO 2026/150382. Das informally refers to the reportedly 8,598-page document as the “Mothership”. Tracing its claimed priority framework back to August 2025 and reportedly backed by eight subsequent international filings totaling more than 10,000 pages of text, the specification reportedly lays out over 300 technical scenarios spanning AI agents, operating systems, cloud infrastructure, digital payments, robotics and telecommunications. All of it is built around a single architectural premise that Das summarizes as “computation is not authority”: an AI model generating a decision, selecting a software tool or holding an authentication token should never by itself trigger an external effect. Instead, proposed actions are held in an inactive state until an independent boundary outside the model verifies fresh permissions and context right before execution.
Within the narrower field of autonomous AI agents, the published filings and companion papers cover several security controls major AI labs are currently working to strengthen. Specific sections describe isolating AI agents inside credential-free sandboxes that hold only dummy handles while a separate host-side proxy swaps in real keys at network exit; splitting an AI model’s tool selection from actual execution on Model Context Protocol servers; restricting writes to an agent’s long-term memory and regulating cumulative data disclosure; gating mobile operating-system actions across iOS and Android; and forcing payment or multi-step enterprise workflows to verify outcomes rather than trusting API success messages.
Rather than leaving those documents sitting quietly in patent registries, Das has spent the past several months distributing derivative specifications across the technical venues engineers and patent examiners monitor most closely, perhaps to realize his technological visions. As of November, he reported maintaining 75 individual Internet-Drafts on the Internet Engineering Task Force Datatracker, paired with formal intellectual-property statements explaining that the broad “Mothership” family is not covered by blanket royalty-capped licensing commitments. He has simultaneously uploaded dozens of papers to Elsevier’s Technical Disclosure Commons, deposited multi-thousand-page specifications on Zenodo, posted machine-readable datasets on Hugging Face, and contributed proposals to the European Commission’s open AI Alliance forum. A busy man indeed. Several of his October technical drafts now have gotten to the point that discovery tags such as “Patent-Examiner, Industry and Platform Security Discovery” are embedded in them.
Das has also begun publishing direct timeline comparisons juxtaposing his mid-2026 WIPO publications with architectures announced later by major technology firms. On 8 September 2026, and in an expanded analysis issued on 5 November, he compared Embodiment 85.B2 of his 16 July WIPO filing with Meta’s agent-security architecture, which Meta unveiled 54 days later on 8 September. Das framed the exercise as an open technical verification rather than an accusation of copying or infringement, inviting security researchers and patent examiners to compare functional correspondences and implementation differences between the two records.
Obviously, there are clear limits to how far an unfunded solo effort can push against Silicon Valley. For example, none of Das’s broad agent-governance claims have been granted as patents, and his Datatracker submissions carry standard notices stating that publication confers no IETF endorsement, standards adoption or confirmation of patent validity. Advancing more than a dozen international filings into separate national patent offices in the United States, Europe and Asia requires substantial legal and filing fees, while his own disclosures note that runnable software references were posted later rather than alongside the initial filings, leaving room for patent examiners to question whether early specifications describe adequately enabled engineering. Patent applications with earlier effective filing dates also remain unaffected by any future antics.
Even so, global patent rules do draw a sharp line between winning a monopoly and spoiling someone else’s. A sufficiently detailed published technical record does not need to mature into a granted patent, win a courtroom verdict or ship as a commercial product to challenge a later applicant’s claim to the same disclosed invention. By placing thousands of pages of dated agent-isolation architectures directly into WIPO, IETF and defensive-publishing indexes, Das may complicate future efforts by companies to patent specific features of autonomous agent control, potentially prompting competitors to draft narrower, implementation-specific claims or address his filings in future disputes.

Leave a Reply