#FactCheck: 2023 Manipur Violence Video Falsely Shared as Indian Army Bombing Minority House

Executive Summary:

Given that AI technologies are evolving at a fast pace in 2024, an AI-oriented phishing attack on a large Indian financial institution illustrated the threats. The documentation of the attack specifics involves the identification of attack techniques, ramifications to the institution, intervention conducted, and resultant effects. The case study also turns to the challenges connected with the development of better protection and sensibilisation of automatized threats. 

Introduction

Due to the advancement in AI technology, its uses in cybercrimes across the world have emerged significant in financial institutions. In this report a serious incident that happened in early 2024 is analysed, according to which a leading Indian bank was hit by a highly complex, highly intelligent AI-supported phishing operation. Attack made use of AI’s innate characteristic of data analysis and data persuasion which led into a severe compromise of the bank’s internal structures.

Background

The chosen financial institution, one of the largest banks in India, had a good background regarding the extremity of its cybersecurity policies. However, these global cyberattacks opened up new threats that AI-based methods posed that earlier forms of security could not entirely counter efficiently. The attackers concentrated on the top managers of the bank because it is evident that controlling such persons gives the option of entering the inner systems as well as financial information. 

Attack Execution

The attackers utilised AI in sending the messages that were an exact look alike of internal messages sent between employees. From Facebook and Twitter content, blog entries, and lastly, LinkedIn connection history and email tenor of the bank’s executives, the AI used to create these emails was highly specific. Some of these emails possessed official formatting, specific internal language, and the CEO’s writing; this made them very realistic. 

 It also used that link in phishing emails that led the users to a pseudo internal portal in an attempt to obtain the login credentials. Due to sophistication, the targeted individuals thought the received emails were genuine, and entered their log in details easily to the bank’s network, thus allowing the attackers access. 

Impact

It caused quite an impact to the bank in every aspect. Numerous executives of the company lost their passwords to the fake emails and compromised several financial databases with information from customer accounts and transactions. The break-in permitted the criminals to cease a number of the financial’s internet services hence disrupting its functions and those of its customers for a number of days. 

 They also suffered a devastating blow to their customer trust because the breach revealed the bank’s weakness against contemporary cyber threats. Apart from managing the immediate operations which dealt with mitigating the breach, the financial institution was also toppling a long-term reputational hit. 

Technical Analysis and Findings

1. The AI techniques that are used in generation of the phishing emails are as follows: 

•  The attack used powerful NLP technology, which was most probably developed using the large-scaled transformer, such as GPT (Generative Pre-trained Transformer). Since these models are learned from large data samples they used the examples of the conversation pieces from social networks, emails and PC language to create quite credible emails. 

 Key Technical Features: 

•  Contextual Understanding: The AI was able to take into account the nature of prior interactions and thus write follow up emails that were perfectly in line with prior discourse. 
•  Style Mimicry: The AI replicated the writing of the CEO given the emails of the CEO and then extrapolated from the data given such elements as the tone, the language, and the format of the signature line. 
•  Adaptive Learning: The AI actively adapted from the mistakes, and feedback to tweak the generated emails for other tries and this made it difficult to detect. 

2. Sophisticated Spear-Phishing Techniques

Unlike ordinary phishing scams, this attack was phishing using spear-phishing where the attackers would directly target specific people using emails. The AI used social engineering techniques that significantly increased the chances of certain individuals replying to certain emails based on algorithms which machine learning furnished. 

 Key Technical Features: 

•  Targeted Data Harvesting: Cyborgs found out the employees of the organisation and targeted messages via the public profiles and messengers were scraped. 
•  Behavioural Analysis: The latest behaviour pattern concerning the users of the social networking sites and other online platforms were used by the AI to forecast the courses of action expected to be taken by the end users such as clicking on the links or opening of the attachments. 
•  Real-Time Adjustments: These are times when it was determined that the response to the phishing email was necessary and the use of AI adjusted the consequent emails’ timing and content. 

3. Advanced Evasion Techniques

The attackers were able to pull off this attack by leveraging AI in their evasion from the normal filters placed in emails. These techniques therefore entailed a modification of the contents of the emails in a manner that would not be easily detected by the spam filters while at the same time preserving the content of the message. 

Key Technical Features: 

•  Dynamic Content Alteration: The AI merely changed the different aspects of the email message slightly to develop several versions of the phishing email that would compromise different algorithms. 
•  Polymorphic Attacks: In this case, polymorphic code was used in the phishing attack which implies that the actual payloads of the links changed frequently, which means that it was difficult for the AV tools to block them as they were perceived as threats. 
•  Phantom Domains: Another tactic employed was that of using AI in generating and disseminating phantom domains, that are actual web sites that appear to be legitimate but are in fact short lived specially created for this phishing attack, adding to the difficulty of detection. 

4. Exploitation of Human Vulnerabilities

This kind of attack’s success was not only in AI but also in the vulnerability of people, trust in familiar language and the tendency to obey authorities. 

Key Technical Features: 

•  Social Engineering: As for the second factor, AI determined specific psychological principles that should be used in order to maximise the chance of the targeted recipients opening the phishing emails, namely the principles of urgency and familiarity. 
•  Multi-Layered Deception: The AI was successfully able to have a two tiered approach of the emails being sent as once the targeted individuals opened the first mail, later the second one by pretext of being a follow up by a genuine company/personality. 

Response

On sighting the breach, the bank’s cybersecurity personnel spring into action to try and limit the fallout. They reported the matter to the Indian Computer Emergency Response Team (CERT-In) to find who originated the attack and how to block any other intrusion. The bank also immediately started taking measures to strengthen its security a bit further, for instance, in filtering emails, and increasing the authentication procedures. 

 Knowing the risks, the bank realised that actions should be taken in order to enhance the cybersecurity level and implement a new wide-scale cybersecurity awareness program. This programme consisted of increasing the awareness of employees about possible AI-phishing in the organisation’s info space and the necessity of checking the sender’s identity beforehand.

Outcome

Despite the fact and evidence that this bank was able to regain its functionality after the attack without critical impacts with regards to its operations, the following issues were raised. Some of the losses that the financial institution reported include losses in form of compensation of the affected customers and costs of implementing measures to enhance the financial institution’s cybersecurity. However, the principle of the incident was significantly critical of the bank as customers and shareholders began to doubt the organisation’s capacity to safeguard information in the modern digital era of advanced artificial intelligence cyber threats. 

 This case depicts the importance for the financial firms to align their security plan in a way that fights the new security threats. The attack is also a message to other organisations in that they are not immune from such analysis attacks with AI and should take proper measures against such threats.

Conclusion

The recent AI-phishing attack on an Indian bank in 2024 is one of the indicators of potential modern attackers’ capabilities. Since the AI technology is still progressing, so are the advances of the cyberattacks. Financial institutions and several other organisations can only go as far as adopting adequate AI-aware cybersecurity solutions for their systems and data. 

 Moreover, this case raises awareness of how important it is to train the employees to be properly prepared to avoid the successful cyberattacks. The organisation’s cybersecurity awareness and secure employee behaviours, as well as practices that enable them to understand and report any likely artificial intelligence offences, helps the organisation to minimise risks from any AI attack.

Recommendations

1. Enhanced AI-Based Defences: Financial institutions should employ AI-driven detection and response products that are capable of mitigating AI-operation-based cyber threats in real-time. 
2.  Employee Training Programs: CYBER SECURITY: All employees should undergo frequent cybersecurity awareness training; here they should be trained on how to identify AI-populated phishing. 
3.  Stricter Authentication Protocols: For more specific accounts, ID and other security procedures should be tight in order to get into sensitive ones. 
4.  Collaboration with CERT-In: Continued engagement and coordination with authorities such as the Indian Computer Emergency Response Team (CERT-In) and other equivalents to constantly monitor new threats and valid recommendations. 
5.  Public Communication Strategies: It is also important to establish effective communication plans to address the customers of the organisations and ensure that they remain trusted even when an organisation is facing a cyber threat. 

Through implementing these, financial institutions have an opportunity for being ready with new threats that come with AI and cyber terrorism on essential financial assets in today’s complex IT environments.

‍

‍

Introduction

In June 2026, Eros Innovation launched Eros Music Worlds, describing itself as the world's first "large cultural music platform". Its first two acts, Jordan and Tanu, are not singers in any conventional sense; they are AI-native personas built from Eros's own film characters, powered by models trained on a licensed corpus of roughly 1.5 trillion cultural tokens drawn from 11,000 films. Alongside them sits a strategic partnership with the family of Mohammed Rafi, the late playback legend, to produce new recordings, a live concert franchise, and a music academy in his name. It's a tidy preview of where AI in music is actually headed: not a novelty generator bolted onto the old industry, but a new kind of media company built around owned IP, licensed data, and characters designed to outlive any single song.

Beyond the demo

The last eighteen months have produced a stream of "firsts". Xania Monet, an AI-voiced R&B act built by Mississippi poet Telisha Jones using Suno, became the first AI artist to chart on a Billboard airplay ranking and was signed by Hallwood Media after a bidding war that reportedly reached $3 million. Breaking Rust, Enlly Blue and Juno Skye followed onto country, rock and Christian charts within weeks. None of this happened in a vacuum: Suno and Udio, the two dominant AI music generators, went from being sued by every major label in mid-2024 to signing licensing settlements with Universal and Warner by late 2025. Voice cloning has run a parallel track from the viral 2023 track "Heart on My Sleeve", which mimicked Drake and The Weeknd without consent, to a wave of Indian cases where AI tools recreated a specific singer's timbre for commercial use. Composition and production tools are now routine studio infrastructure rather than curiosities; the interesting frontier has shifted to multilingual adaptation and catalogue reanimation. Eros has promised 34-language localisation for its AI artists, a scale no dubbing studio could match manually. A related, if less music-specific, signal came from Collective Artists Network's AI-powered Mahabharat: Ek Dharmayudh, produced with Prasar Bharati. AI tools handled pre-visualisation, multi-language dubbing sync and sound design, while composers and voice directors retained control of tone and emotion, a template other Indian studios are likely to copy for scoring and localising large-scale cultural content cheaply.

Where the money actually is?

The commercial case for labels isn't "AI makes music at scale", because scale alone isn't value. Deezer, the only major platform publishing detection data, reported in July 2026 that AI-generated tracks had crossed 50% of daily uploads, around 90,000 tracks a day, up from 10,000 in January 2025. Yet those tracks account for only 1–3% of actual streams, and Deezer flags 85% of that listening as fraudulent bot activity. Generating music has become nearly free; getting anyone to listen hasn't. The real opportunity for rights holders lies elsewhere: licensing catalogues into controlled ecosystems (Universal's 2026 platform with Udio, Warner's deals with both Udio and Suno), reviving legacy artists for new audiences (Eros-Rafi), and using AI for the unglamorous, high-volume work of localisation, sync placement and short-form content that never justified full production budgets before. AI-native performers also travel well into gaming, virtual concerts and brand collaboration formats built for characters rather than people, where a synthetic act never ages, never cancels a tour and can be licensed into a dozen campaigns simultaneously. That is a genuinely new revenue line, but it's a narrow one; it monetises IP ownership, not music generation itself. IFPI's Global Music Report 2026 shows the underlying business is healthy: revenues reached $31.7 billion, up 6.4% — but that growth is still driven overwhelmingly by human catalogues and paid subscriptions, not AI-generated volumes.

The legal fault lines

This is where India's position is genuinely instructive and unresolved. The Copyright Act, 1957, defines the author of a "computer-generated" work as the person who causes it to be created (language added in 1994 for a world of human-directed software, not autonomous generative models), and Indian courts haven't yet settled how far it stretches. What's clearer is the industry's stance on training data: The Indian music industry body, alongside T-Series, Saregama and Sony Music, has sought to join the ANI Media suit against OpenAI in the Delhi High Court, arguing that scraping sound recordings to train models without a licence breaches copyright, a case they say is "crucial for the entire music industry in India, and even worldwide".

Personality rights have moved faster than legislation. In the absence of a dedicated statute, Indian courts have built protection case by case: the Delhi High Court's 2023 order shielding Anil Kapoor's voice and likeness from AI misuse and the Bombay High Court's 2024 ruling in Arijit Singh v Codible Ventures, India's first AI voice-cloning judgement, which granted an injunction spanning every medium, explicitly including the metaverse and generative-AI tools, and rejected fair-use and parody defences outright. Similar orders have since followed for Amitabh Bachchan, Aishwarya Rai Bachchan and others. The pattern mirrors, but predates, statutory responses abroad: Tennessee's 2024 ELVIS Act was the first US law to protect voice explicitly as property, and the federal NO FAKES Act remains pending. The US Copyright Office's January 2025 report reaffirmed that human authorship is the bedrock of copyrightability, denying protection to fully AI-generated output while leaving substantially human-directed hybrid works to case-by-case review. Two regimes are converging on the same instinct that consent and human contribution must anchor ownership through different legal tools: India via judge-made personality rights and the US via statute and Copyright Office guidance.

Who bears the risk?

The disruption isn't evenly spread. Session singers, jingle artists, dubbing performers and early-career composers, the layer of the industry that supplies craft rather than fame, are most exposed since their work is precisely what generative tools now approximate cheaply. A PRS for Music survey of over 2,600 creators found 79% worried about AI competing directly with their output and 76% expecting it to hurt their livelihoods. CISAC has projected AI could erode musicians' incomes by nearly a quarter by 2028. Separately, roughly a fifth of professional voice actors surveyed say they've already lost work to a synthetic voice. For streaming platforms, the risk is structural: as AI supply floods catalogues, royalty pools shared across all streams get diluted, discoverability worsens, and fraud bot streams on AI tracks are designed purely to harvest royalties, which becomes a live threat to the economics everyone else depends on.

Recommendations

None of this points to AI simply replacing musicians, nor to it staying a background production tool. A third category is emerging AI-native "characters" with continuity, backstory and owned IP, closer to franchise properties than bands, which is exactly the model Eros and Hallwood Media are betting on. The deeper shift underneath it is economic: once the marginal cost of producing a new track approaches zero, the constraint on the industry stops being supplied and becomes attention. Whether AI-native artists become a durable category depends on questions still unresolved, training data legality chief among them, and on whether audiences keep listening once the novelty fades. Deezer's own research found 97% of listeners can't reliably distinguish AI from human music, yet 52% still believe fully AI-generated songs shouldn't sit on the same charts as human ones. Provenance matters even when perception can't detect it.

Conclusion

Scarcity in the age of AI will not be creative output but genuine human connection. As AI makes music production abundant and inexpensive, the qualities it cannot replicate authenticity, lived experience, artistic identity, and trust – become increasingly valuable. Audiences will continue to seek creators with compelling stories and meaningful live experiences. In an era of infinite synthetic content, originality is no longer the premium; credibility, human presence, and emotional connection are.

Sources

1. https://www.livemint.com/industry/media/ai-music-ai-native-artists-eros-music-worlds-eros-innovation-copyright-intellectual-property-music-industry-11785053693997.html
2. https://www.facebook.com/KJSIMofficial/posts/the-mint-featured-insights-from-dr-alka-agarwal-assistant-professor-general-mana/1375301184595168/
3. https://www.linkedin.com/posts/livemint_how-ai-brought-the-mahabharat-to-life-collective-activity-7389643662193999873-6BKG
4. https://musically.com/2026/07/02/ai-powered-platform-eros-music-worlds-is-turning-film-characters-into-virtual-singers/
5. https://www.business-standard.com/companies/news/bollywood-t-series-saregama-sony-music-copyright-lawsuit-openai-delhi-hc-125021400876_1.html 

‍

Introduction

In a world where Artificial Intelligence (AI) is already changing the creation and consumption of content at a breathtaking pace, distinguishing between genuine media and false or doctored content is a serious issue of international concern. AI-generated content in the form of deepfakes, synthetic text and photorealistic images is being used to disseminate misinformation, shape public opinion and commit fraud. As a response, governments, tech companies and regulatory bodies are exploring ‘watermarking’ as a key mechanism to promote transparency and accountability in AI-generated media. Watermarking embeds identifiable information into content to indicate its artificial origin. 

Government Strategies Worldwide

Governments worldwide have pursued different strategies to address AI-generated media through watermarking standards. In the US, President Biden's 2023 Executive Order on AI directed the Department of Commerce and the National Institute of Standards and Technology (NIST) to establish clear guidelines for digital watermarking of AI-generated content. This action puts a big responsibility on large technology firms to put identifiers in media produced by generative models. These identifiers should help fight misinformation and address digital trust.

The European Union, in its Artificial Intelligence Act of 2024, requires AI-generated content to be labelled. Article 50 of the Act specifically demands that developers indicate whenever users engage with synthetic content. In addition, the EU is a proponent of the Coalition for Content Provenance and Authenticity (C2PA), an organisation that produces secure metadata standards to track the origin and changes of digital content.

India is currently in the process of developing policy frameworks to address AI and synthetic content, guided by judicial decisions that are helping shape the approach. In 2024, the Delhi High Court directed the central government to appoint members for a committee responsible for regulating deepfakes. Such moves indicate the government's willingness to regulate AI-generated content. 

China, has already implemented mandatory watermarking on all deep synthesis content. Digital identifiers must be embedded in AI media by service providers, and China is one of the first countries to adopt stern watermarking legislation.

Understanding the Technical Feasibility

Watermarking AI media means inserting recognisable markers into digital material. They can be perceptible, such as logos or overlays or imperceptible, such as cryptographic tags or metadata. Sophisticated methods such as Google's SynthID apply imperceptible pixel-level changes that remain intact against standard image manipulation such as resizing or compression. Likewise, C2PA metadata standards enable the user to track the source and provenance of an item of content.

Nonetheless, watermarking is not an infallible process. Most watermarking methods are susceptible to tampering. Aforementioned adversaries with expertise, for instance, can use cropping editing or AI software to delete visible watermarks or remove metadata. Further, the absence of interoperability between different watermarking systems and platforms hampers their effectiveness. Scalability is also an issue enacting and authenticating watermarks for billions of units of online content necessitates huge computational efforts and routine policy enforcement across platforms. Scientists are currently working on solutions such as blockchain-based content authentication and zero-knowledge watermarking, which maintain authenticity without sacrificing privacy. These new techniques have potential for overcoming technical deficiencies and making watermarking more secure.

Challenges in Enforcement

Though increasing agreement exists for watermarking, implementation of such policies is still a major issue. Jurisdictional constraints prevent enforceability globally. A watermarking policy within one nation might not extend to content created or stored in another, particularly across decentralised or anonymous domains. This creates an exigency for international coordination and the development of worldwide digital trust standards. While it is a welcome step that platforms like Meta, YouTube, and TikTok have begun flagging AI-generated content, there remains a pressing need for a standardised policy that ensures consistency and accountability across all platforms. Voluntary compliance alone is insufficient without clear global mandates.

User literacy is also a significant hurdle. Even when content is properly watermarked, users might not see or comprehend its meaning. This aligns with issues of dealing with misinformation, wherein it's not sufficient just to mark off fake content, users need to be taught how to think critically about the information they're using. Public education campaigns, digital media literacy and embedding watermarking labels within user-friendly UI elements are necessary to ensure this technology is actually effective.

Balancing Privacy and Transparency

While watermarking serves to achieve digital transparency, it also presents privacy issues. In certain instances, watermarking might necessitate the embedding of metadata that will disclose the source or identity of the content producer. This threatens journalists, whistleblowers, activists, and artists utilising AI tools for creative or informative reasons. Governments have a responsibility to ensure that watermarking norms do not violate freedom of expression or facilitate surveillance. The solution is to achieve a balance by employing privacy-protection watermarking strategies that verify the origin of the content without revealing personally identifiable data. "Zero-knowledge proofs" in cryptography may assist in creating watermarking systems that guarantee authentication without undermining user anonymity.

On the transparency side, watermarking can be an effective antidote to misinformation and manipulation. For example, during the COVID-19 crisis, misinformation spread by AI on vaccines, treatments and public health interventions caused widespread impact on public behaviour and policy uptake. Watermarked content would have helped distinguish between authentic sources and manipulated media and protected public health efforts accordingly.

Best Practices and Emerging Solutions

Several programs and frameworks are at the forefront of watermarking norms. Adobe, Microsoft and others' collaborative C2PA framework puts tamper-proof metadata into images and videos, enabling complete traceability of content origin. SynthID from Google is already implemented on its Imagen text-to-image model and secretly watermarks images generated by AI without any susceptibility to tampering. The Partnership on AI (PAI) is also taking a leadership role by building out ethical standards for synthetic content, including standards around provenance and watermarking. These frameworks become guides for governments seeking to introduce equitable, effective policies. In addition, India's new legal mechanisms on misinformation and deepfake regulation present a timely point to integrate watermarking standards consistent with global practices while safeguarding civil liberties.

Conclusion

Watermarking regulations for synthetic media content are an essential step toward creating a safer and more credible digital world. As artificial media becomes increasingly indistinguishable from authentic content, the demand for transparency, origin, and responsibility increases. Governments, platforms, and civil society organisations will have to collaborate to deploy watermarking mechanisms that are technically feasible, compliant and privacy-friendly. India is especially at a turning point, with courts calling for action and regulatory agencies starting to take on the challenge. Empowering themselves with global lessons, applying best-in-class watermarking platforms and promoting public awareness can enable the nation to acquire a level of resilience against digital deception. 

References

1. https://artificialintelligenceact.eu/
2. https://www.cyberpeace.org/resources/blogs/delhi-high-court-directs-centre-to-nominate-members-for-deepfake-committee
3.  https://c2pa.org
4. https://www.cyberpeace.org/resources/blogs/misinformations-impact-on-public-health-policy-decisions
5.  https://deepmind.google/technologies/synthid/
6. https://www.imatag.com/blog/china-regulates-ai-generated-content-towards-a-new-global-standard-for-transparency

‍