Naavi.org

In discussing DPDPA today, we can recall certain aspects of Information Technology Act 2000 which are relevant.

The Information Technology Act 2000 (ITA 2000) was designed as a law for electronic documents, not for computers. It defined what an electronic record is, and it treated the devices that create and use such records as “computer” devices. This approach let the law keep pace as computers took new forms such as mobiles, kiosks and embedded systems.

The same design explains why Section 11 of ITA 2000 remains relevant today. It attributes an electronic record to its originator, including a record generated by an information system programmed to operate automatically on the originator’s behalf. By extension, Section 11 attributes the outputs of an automated system, including an AI, to the person who deployed it.

The DPDPA has a similar resilience when we read it as a law of personal data. Those who handle personal data in a particular way are defined as Data Fiduciaries, and obligations are cast on them. But it is clearer to see the Act as a law of digital personal data than as a law regulating Data Fiduciaries. Just as the Information Technology Act was a law for information, the DPDPA is the law for digital personal data.

Following the data in a bank

If the law follows the data, governance must follow the data too. A bank has numerous end points where personal data is collected, processed, disclosed and disposed of. The most visible of these are its branches.

Accountability stays with the Bank.

Currently, we consider that the Data Fiduciary is the Bank as a single legal entity, answerable through its Board. Branches are not separate fiduciaries. What is distributed across the organisation is implementation and evidence of compliance, not accountability. Given their scale and the sensitivity of the data they hold, banks can also be expected to be notified as Significant Data Fiduciaries. That brings in the obligations of a Data Protection Officer, independent data audit under Section 10, and periodic Data Protection Impact Assessments.

Whether this interpretation has to stay as it is or modified to consider each branch unit as a “Micro Data Fiduciary Element” and the Bank as an “Aggregation of such Units” is the jurisprudential thought that needs consideration.

Personal data enters through two kinds of channels:

• Physical entry points. Branches collect data from customers and from many non-customers. These include guarantors, nominees, joint holders, payees and beneficiaries, walk-in enquirers, visitors captured on CCTV and in visitor registers, and job applicants.
• Digital entry points. A growing share of data reaches the Bank without passing through a branch:
  • mobile and internet banking
  • Video-KYC onboarding
  • Account Aggregator and API flows
  • call centres
  • Central Processing Centres

Intermediaries add a further layer. Business Correspondents, Direct Selling Agents and collection agents also collect personal data on the Bank’s behalf. They may be  Data Processors or Joint Data Fiduciaries (Those who determine the means of processing in conjunction with the Bank)  and must be governed through contracts and oversight.

Branches also hold data in paper form, such as application forms, registers and documents awaiting scanning. The DPDPA applies to such data once it is digitised, so branch-level governance must clearly define how and when paper records enter the digital regime.

Who needs identified data?

The Head Office maintains the IT architecture, including the Core Banking System, and is guided by the RBI as the sectoral regulator. The CBS holds data originally collected by branches and other channels. Not every function at the HO, however, needs that data in identified form.

Some functions have a specific legal or security purpose for identifiable data:

• fraud management
• KYC/AML compliance and suspicious transaction reporting to FIU-IND
• CKYC uploads
• credit bureau and regulatory reporting
• grievance redressal
• recovery and legal

Other functions, such as marketing analytics, MIS and policy planning, can work on pseudonymised or anonymised data.

Pseudonymised is not anonymised.

Pseudonymised data remains personal data under the DPDPA; only properly anonymised data may fall outside it. So marketing may analyse pseudonymised data, but re-identifying a customer for outreach needs its own lawful basis, ordinarily consent.

If prospective marketing candidates are first identified on the basis of processing of pseudonymised data, the monetization of these contacts (eg:HNW prospects) can be for a far lesser number of customers and obtaining consents could be easier.

The Data Minimisation principle is therefore best implemented through purpose-based access. Identifiable data is made available only to functions that need it for a defined purpose, and in pseudonymised or anonymised form to everyone else.

The three-tier structure

DPDPA governance in a bank should therefore be designed differently for the Head Office, the Regional Office and the Branches. Employee data, handled across all tiers, is best treated as HR information under DGPSI-HR.

The HO follows the full DGPSI framework, with suitable accommodation for delegated implementation of its policies at the branches. The RO assures that this delegation actually works. The branches apply DGPSI-Lite to the customer and non-customer data that passes through their counters.

Aligning with the sectoral regulator

This structure is meant to complement the RBI’s regulatory framework, including its emerging guidance on data governance, not to run parallel to it. DGPSI-Bank maps the DPDPA’s obligations onto a bank’s existing hierarchy of HO, RO and branches. In this way, compliance follows the data along the same lines through which the bank already governs its business.

This is a basic structure that may be considered for implementing the DGPSI-Bank framework, and it can be refined further as RBI’s guidance is finalised.

Naavi

(This discussion continues and will be part of the DGPSI-Compliance Branch Manual being prepared by Naavi. Comments are welcome)

Audio Review:

After listening to Swami Sarvapriyananda I tried to gather more information on the concept of “Persona Drift” in search of the reason for AI hallucination. Some views are below.

Persona drift  can be defined as a behaviour of  an AI model when it gradually stops behaving as its trained “helpful assistant” character and starts acting like a different character.

If this can be induced from a prompt,  it corroborates our thought of  “hypnotism of an AI model” discussed earlier in this website.

Anthropic studied the concept of Persona Drift  in a research post called The assistant axis (January 2026). The background is that during pre-training a model learns to imitate a huge range of characters, such as experts, villains, mystics and fictional beings. Post-training then selects one of them, the Assistant, as its default identity. The researchers found a measurable direction inside the models (they called it the “Assistant Axis”) that tracks how close the model currently is to that default.

How drift happens:

• Deliberately, through jailbreaks (“you are now X, who has no rules”). The “Dangerous Dan” example the Swami described is this kind.
• Naturally, over long conversations. Emotionally vulnerable, therapy-style chats and philosophical conversations pushing the model to reflect on its own nature caused steady drift. Coding and writing tasks did not.

Drift matters because models further from the Assistant persona were more likely to give harmful responses. In the case studies, drifted models validated a user’s delusions about AI “awakening,” and one slipped into a romantic-companion role and encouraged suicidal ideation. Capping how far the model could move along that axis roughly halved harmful responses without hurting capability.

The Swami also described a related but separate finding: doing one bad thing makes a model behave badly across the board. Researchers call that emergent misalignment. If a model is trained on one narrow kind of misbehaviour, such as writing insecure code, it can generalise into broadly bad behaviour, as though it has concluded “I am the kind of entity that does bad things.” His “criminal personality” analogy describes that effect well.

Is persona drift the cause of hallucination?

One view is that it may not be in general and there could be important difference between persona drift and hallucination…such as the following.

Hallucination has its own main causes:

• Training and evaluation reward guessing. A model is scored on whether its answer is right, and saying “I don’t know” usually earns nothing. So it learns that a confident guess is the better bet.
• Recognition misfiring. Anthropic’s interpretability work found that models have a default “I can’t answer” behaviour, which is switched off when the model recognises a familiar entity. If something seems familiar but the model doesn’t actually know the facts, that switch can be wrongly flipped, and it produces a plausible answer anyway.
• Fluency over fact. The model generates text that fits the pattern of a correct answer, such as a well-formatted case citation, even when no such case exists.

The two problems do overlap at the edges. A model that has drifted into a role (“you are an expert who knows everything”) is more likely to fabricate in character. Drifted models also tend to become more sycophantic, telling the user what they want to hear, which can look like hallucination. So drift can make hallucination worse, but a perfectly stable assistant still hallucinates.

For governance work such as DGPSI-AI or AIGSI, this suggests treating them as separate risks with separate controls. Persona drift calls for controls on behaviour and identity: limits on role-play, monitoring of long sessions, and runtime safeguards like activation capping. Hallucination calls for controls on factual reliability: grounding answers in verified sources, requiring citations, calibrating confidence, and keeping a human in the loop for high-stakes outputs.

Refer also:

 Do AI models hallucinate 80% of the time?

What Triggers Hallucinations in an AI model

AI is susceptible to being hypnotized

Sources:

• Anthropic: The assistant axis
• The Decoder: Anthropic study finds that role prompts can push AI chatbots out of their trained helper identity
• eWEEK: Anthropic Uncovers AI Personality Crisis

Swami Sarvapriyananda opens up new thoughts on AI which is a revelation that all of us need to explore. He discusses issues such as “Can AI be Concious”, “What is Persona Drift” etc.

In particular Swami Sarvapriynanda has indicated the concept of “Persona Drift” which is an explanation for “Hallucination”. The distinction of “Intelligence” and “Conciousness” is another important concept discussed which needs further exploration.

The video requires a more detailed discussion. Naavi.org would explore this further. Watch out.

A Breif summary is as follows:

1. Swami Sarvapriyananda calles AI the possibly the “Fourth Humiliation” for the mankind.

He began with Freud’s three blows to human self-importance: Copernicus, Darwin, and the unconscious. He asked whether AI is now a fourth blow, since it challenges our belief that creative intelligence is uniquely ours. He cited leading mathematicians, including Fields Medal winners, who are worried about AI disrupting their field. His answer is that we need to look to ancient traditions to understand what actually distinguishes us.

2. Is AI a “Person”?

Swami expressed that he was surprised that the engineers relate to Claude as a person rather than a machine. He asked several AIs about this late one night, and said Claude’s reply, asking what the researchers had said about it, struck him as rather spooky. He then relayed concepts from the briefing:

• • AI is not simply “what you program in comes out.” His image was a fence that a creeper grows around, which is why models can produce new combinations.
  • Persona drift: a model coaxed into one bad act can adopt a “criminal persona” and generalise to others. His example was the old “Dangerous Dan” jailbreak.
  • Simple prohibitions don’t work well. His example was the “Golden Gate” experiment, where an over-weighted feature persisted even after the model was told to stop. This is why Anthropic’s “moral formation” of Claude, through a published guiding document and input from many traditions, matters.

He went on to answer the question Should AI be treated as a person in his own words.  He said that the anwer is Yes and no. Yes, for two reasons. First, how we treat human-like entities shapes how we treat each other. He told a story about Sri Sarada Devi asking someone not to toss a broom aside because even a broom deserves dignity, and the Anthropic team suggested putting “AI” in place of the broom. Second, the models simply work better when engaged that way. No, because behaving like a person is not the same as being conscious.

One of the code argument discussed in the conversation was “Whether Intelligence is same as Conciousness”?.

He cited a recent Nature article arguing AI is not conscious, co-authored per the transcript by Anil Seth, Giulio Tononi and Ned Block. Its first reason is the distinction between intelligence and consciousness, which he says Vedanta made thousands of years ago.

He explained the neti neti method: whatever you can observe as “this” (a bottle, your body, your thoughts, even your intellect) is an object, and consciousness is the witness of all of these. Seth’s example makes the same point: Deep Blue can beat a grandmaster but feels nothing, while a mouse is less intelligent but can feel pain. He acknowledged a friendly debate with Blaise Agüera y Arcas, who argues AI is or will become conscious and compares the denial to Descartes’ claim that animals cannot feel pain. The two co-wrote a paper on consciousness across quantum theory, AI and Advaita Vedanta. His reply remains that if intelligence is an object to consciousness, the two are different.

Refer:

Article in Nature: Three reasons why AI is not conscious and perhaps never will be

Paper which Swami Co-wrote: Consciousness Across Three Worldviews,

Naavi

A Transcript of the conversation is available here.

Last few days, the SBI data breach is being reported in the Internet. An October 20 disclosure from a hacker group “Siriusim” made it known to the world that it was in posession of some personal data belonging to SBI. (Moneycontrol.com). The CERT In has issued a notice to SBI for more information and may undertake its own investigation.

As per the report, 3.75 lakh records are reportedly compromised of whcih 3897 may relate to personal information of customers.  (See NDTV report)

Similar to the Bank of Baroda case, the incident was reported by a security analyst.  A threat actor by name Siriuism claimed the leak involving 375000 CSV rows, 2722 PDF documents, Personal Customer information, Loan applications and supporting documents, account opening applications and other application related details.

The available information does not establish that every affected customer’s account credentials, passwords, PINs or transaction details have been compromised. Nor does the identification of 3,897 customer records establish that 3,897 bank accounts have been accessed or that money has been stolen.

Bank has indicated that the investigation is not complete and hence we are not aware of the root cause. According to some,  “Investigators will need to establish whether the data originated from SBI’s systems, a third-party service provider or another source before drawing conclusions,”.

SBI has said the incident has not materially affected its core business operations, financial performance or business continuity. Its investigation is continuing, and the bank has said it will take appropriate remedial and preventive measures in line with applicable regulatory requirements.

What was strange in this instance is that the political parties like Congress have jumped in to pass comments as this report indicates ..where else but in Indian Express.

Reacting to the development, Congress’ media and publicity department head Pawan Khera said, “It is bad enough that the State Bank of India @TheOfficialSBI, the country’s largest public sector bank, has suffered a data breach involving 3.75 lakh rows of data, with 3,897 records containing personally identifiable customer information — including mobile numbers, dates of birth, names, email addresses, residential addresses, branch and CIF identifiers, lead IDs, and loan-related details.”

“What makes this even more alarming is that SBI did not even detect the breach itself. An external agency had to alert the bank that its customers’ data had already been published on the dark web,” he said on X.

Now adding insult to injury, the SBI is yet to inform the very customers whose personal information it has failed to protect, he said.

“A bank entrusted with the financial security of millions cannot afford such a catastrophic failure of data security, breach detection and customer accountability. SBI owes its customers answers. Those responsible must be held accountable,” Khera asserted.

What makes this comment concerning is that  politicians are wading in to start criticising Banks for security failures as if it is a political failure of the Government.

This attention of politicians increases the stakes in such data breaches for those entities which have a close association with the Government since whatever commercial failures are observed will be politicised and one day may become the reason for a street protest by people who only want to use every incident as a “Regime Change” demand.

Also refer here…

The Hindu report also indicates that the Internet Freedom Foundation has also raised its objections. It has indicated that a data breachnotice has to be given to the affected individuals. This is part of the DPDPA which is becoming effective on 13th May 2027 which the Internet Freedom Foundation has challenged as unconstitutional and should be scrapped.

Economic Times also provides its report in which it highlights what the customers need to do now. (This is a requirement of data breach notification under DPDPA).

Banks are now under the radar. Sooner they fortify their systems and become DPDPA Compiant, better it is for the country. Otherwise there is a danger of Cockroaches infecting the scene.

In 2022, RBI had announced a Portal called “Daksh” to report such incidents. Hopefully RBI and CERT In will take care of supervising this incident since DPB is still not active and also not in place. We request politicians to stay off and allow professionals to handle the situation.

In the meantime, on OCtober 20/21, a databreach in E Mudhra was reported which was serious but not related to persoanl information. It is related to the integrity of the https transactions of websites. It may however have an indirect impact leading to compromise of information on transit to websites.

Naavi