What’s Your New Year's Resolution?

As AI language models become more powerful, they are also becoming more prone to errors. One increasingly prominent issue is AI hallucinations, instances where models generate outputs that are factually incorrect, nonsensical, or entirely fabricated, yet present them with complete confidence. Recently, ChatGPT released two new models—o3 and o4-mini, which differ from earlier versions as they focus more on step-by-step reasoning rather than simple text prediction. With the growing reliance on chatbots and generative models for everything from news summaries to legal advice, this phenomenon poses a serious threat to public trust, information accuracy, and decision-making.

What Are AI Hallucinations?

AI hallucinations occur when a model invents facts, misattributes quotes, or cites nonexistent sources. This is not a bug but a side effect of how Large Language Models (LLMs) work, and it is only the probability that can be reduced, not their occurrence altogether. Trained on vast internet data, these models predict what word is likely to come next in a sequence. They have no true understanding of the world or facts, they simulate reasoning based on statistical patterns in text. What is alarming is that the newer and more advanced models are producing more hallucinations, not fewer. seemingly counterintuitive. This has been prevalent reasoning-based models, which generate answers step-by-step in a chain-of-thought style. While this can improve performance on complex tasks, it also opens more room for errors at each step, especially when no factual retrieval or grounding is involved.

As per reports shared on TechCrunch, it mentioned that when users asked AI models for short answers, hallucinations increased by up to 30%. And a study published in eWeek found that ChatGPT hallucinated in 40% of tests involving domain-specific queries, such as medical and legal questions. This was not, however, limited to this particular Large Language Model, but also similar ones like DeepSeek. Even more concerning are hallucinations in multimodal models like those used for deepfakes. Forbes reports that some of these models produce synthetic media that not only look real but are also capable of contributing to fabricated narratives, raising the stakes for the spread of misinformation during elections, crises, and other instances.

It is also notable that AI models are continually improving with each version, focusing on reducing hallucinations and enhancing accuracy. New features, such as providing source links and citations, are being implemented to increase transparency and reliability in responses.

The Misinformation Dilemma

The rise of AI-generated hallucinations exacerbates the already severe problem of online misinformation. Hallucinated content can quickly spread across social platforms, get scraped into training datasets, and re-emerge in new generations of models, creating a dangerous feedback loop. However, it helps that the developers are already aware of such instances and are actively charting out ways in which we can reduce the probability of this error. Some of them are:

1. Retrieval-Augmented Generation (RAG): Instead of relying purely on a model’s internal knowledge, RAG allows the model to “look up” information from external databases or trusted sources during the generation process. This can significantly reduce hallucination rates by anchoring responses in verifiable data.
2. Use of smaller, more specialised language models: Lightweight models fine-tuned on specific domains, such as medical records or legal texts. They tend to hallucinate less because their scope is limited and better curated.

Furthermore, transparency mechanisms such as source citation, model disclaimers, and user feedback loops can help mitigate the impact of hallucinations. For instance, when a model generates a response, linking back to its source allows users to verify the claims made.

Conclusion

AI hallucinations are an intrinsic part of how generative models function today, and such a side-effect would continue to occur until foundational changes are made in how models are trained and deployed. For the time being, developers, companies, and users must approach AI-generated content with caution. LLMs are, fundamentally, word predictors, brilliant but fallible. Recognising their limitations is the first step in navigating the misinformation dilemma they pose.

References

• https://www.eweek.com/news/ai-hallucinations-increase/
• https://www.resilience.org/stories/2025-05-11/better-ai-has-more-hallucinations/ 
• https://www.ekathimerini.com/nytimes/1269076/ai-is-getting-more-powerful-but-its-hallucinations-are-getting-worse/ 
• https://techcrunch.com/2025/05/08/asking-chatbots-for-short-answers-can-increase-hallucinations-study-finds/ 
• https://en.as.com/latest_news/is-chatgpt-having-robot-dreams-ai-is-hallucinating-and-producing-incorrect-information-and-experts-dont-know-why-n/
• https://www.newscientist.com/article/2479545-ai-hallucinations-are-getting-worse-and-theyre-here-to-stay/
• https://www.forbes.com/sites/conormurray/2025/05/06/why-ai-hallucinations-are-worse-than-ever/ 
• https://towardsdatascience.com/how-i-deal-with-hallucinations-at-an-ai-startup-9fc4121295cc/ 
• https://www.informationweek.com/machine-learning-ai/getting-a-handle-on-ai-hallucinations 

‍

Recent Incidents:

Recent reports are revealing a significant security threat linked to a new infostealer based malware campaign known to solely target gaming accounts. This attack has affected users of Activision and other gaming websites. The sophisticated software has captured millions of login credentials, notably from the cheats and players. The officials at Activision Blizzard, an American video game holding company, are still investigating the matter and collaborating with cheated developers to minimize the impact and inform the accounts’ residents of appropriate safety measures.

Overview:

Infostealer, also known as information stealer, is a type of malware designed in the form of a Trojan virus for stealing private data from the infected system. It can have a variety of incarnations and collect user data of various types such as browser history, passwords, credit card numbers, and login details and credentials to social media, gaming platforms, bank accounts, and other websites. Bad actors use the log obtained as a result of the collection of personal records to access the victim’s financial accounts, appropriate the victim’s online identity, and perform fraudulent actions on behalf of the victim.

Modus Operandi:

• Infosteale­r is a malicious program created to illegally obtain pe­ople's login details, like use­rnames and passwords. Its goal is to enable cybe­rattacks, sell on dark web markets, or pursue­ malicious aims.

• This malware targets both personal de­vices and corporate systems. It spre­ads through methods like phishing emails, harmful we­bsites, and infected public site­s.

• Once inside a device­, Infostealer secre­tly gathers sensitive data like­ passwords, account details, and personal information. It's designe­d to infiltrate systems being undete­cted. The stolen cre­dentials are compiled into datalogs. The­se logs are then sold ille­gally on dark web marketplaces for profit.

Analysis: 

‍

‍

Basic properties:

• MD5: 06f53d457c530635b34aef0f04c59c7d
• SHA-1: 7e30c3aee2e4398ddd860d962e787e1261be38fb
• SHA-256: aeecc65ac8f0f6e10e95a898b60b43bf6ba9e2c0f92161956b1725d68482721d
• Vhash: 145076655d155515755az4e?z4
• Authentihash: 65b5ecd5bca01a9a4bf60ea4b88727e9e0c16b502221d5565ae8113f9ad2f878
• Imphash: f4a69846ab44cc1bedeea23e3b680256
• Rich PE header hash: ba3da6e3c461234831bf6d4a6d8c8bff 
• SSDEEP: 6144:YcdXHqXTdlR/YXA6eV3E9MsnhMuO7ZStApGJiZcX8aVEKn3js7/FQAMyzSzdyBk8:YIKXd/UgGXS5U+SzdjTnE3V
• TLSH:T1E1B4CF8E679653EAC472823DCC232595E364FB009267875AC25702D3EFBB3D56C29F90
• File type: Win32 DLL executable windows win32 pepe dll
• Magic: PE32+ executable (DLL) (GUI) x86-64, for MS Windows
• File size: 483.50 KB (495104 bytes)

Additional Hash Files:

• 160389696ed7f37f164f1947eda00830
• 229a758e232aeb49196c862655797e12
• 23e4ac5e7db3d5a898ea32d27e8b7661
• 3440cced6ec7ab38c6892a17fd368cf8
• 36d7da7306241979b17ca14a6c060b92
• 38d2264ff74123f3113f8617fabc49f6
• 3c5c693ba9b161fa1c1c67390ff22c96
• 3e0fe537124e6154233aec156652a675
• 4571090142554923f9a248cb9716a1ae
• 4e63f63074eb85e722b7795ec78aeaa3
• 63dd2d927adce034879b114d209b23de
• 642aa70b188eb7e76273130246419f1d
• 6ab9c636fb721e00b00098b476c49d19
• 71b4de8b5a1c5a973d8c23a20469d4ec
• 736ce04f4c8f92bda327c69bb55ed2fc
• 7acfddc5dfd745cc310e6919513a4158
• 7d96d4b8548693077f79bc18b0f9ef21
• 8737c4dc92bd72805b8eaf9f0ddcc696
• 9b9ff0d65523923a70acc5b24de1921f
• 9f7c1fffd565cb475bbe963aafab77ff

Indicators of Compromise:

• Unusual Outbound Network Traffic: An increase in odd or questionable outbound network traffic may be a sign that infostealer malware has accessed more data.

• Anomalies in Privileged User Account Activity: Unusual behavior or illegal access are two examples of irregular actions that might indicate a breach in privileged user accounts.

• Suspicious Registry or System File Changes: Infostealer malware may be trying to alter system settings if there are any unexpected changes to system files, registry settings, or configurations.

• Unusual  DNS queries: When communicating with command and control servers or rerouting traffic, infostealer malware may produce strange DNS queries.

• Unexpected System Patching: Unexpected or unauthorized system patching by unidentified parties may indicate that infostealer malware has compromised the system and is trying to hide its footprint or become persistent. 

• Phishing emails and social engineering attempts:  It is a  popular strategy employed by cybercriminals to get confidential data or implant malicious software. To avoid compromise, it is crucial to be wary of dubious communications and attempts of social engineering.

Recommendations:

• Be Vigilant: In today's digital world, many cybercrime­s threaten online safe­ty, Phishing tricks, fake web pages, and bad links pose­ real dangers. Carefully che­ck email sources. Examine we­bsites closely. Use top se­curity programs. Follow safe browsing rules. Update software­ often. Share safety tips. The­se steps reduce­ risks. They help kee­p your online presence­ secure.

• Regular use of Anti-Virus Software to detect the threats: Antivirus tools are vital for finding and stopping cybe­r threats. These programs use­ signature detection and be­havior analysis to identify known malicious code and suspicious activities. Updating virus de­finitions and software-patches regularly, improve­s their ability to detect ne­w threats. This helps maintain system se­curity and data integrity.

• Provide security related training to the employees and common employees: One should learn Cybe­rsecurity and the best practice­s in order to keep the­ office safe. Common workers will ge­t lessons on spotting risks and responding well, cre­ating an environment of caution.

• Keep changing passwords: Passwords should be changed fre­quently for better se­curity. Rotating passwords often makes it harder for cybe­r criminals to compromise and make it happen or confidential data to be­ stolen. This practice keeps intruders out and shie­lds sensitive intel.

Conclusion:

To conclude, to reduce the impact and including the safety measures, further investigations and collaboration are already in the pipeline regarding the recent malicious software that takes advantage of gamers and has stated that about millions of credentials users have been compromised. To protect sensitive data, continued usage of antivirus software, use of trusted materials  and password changes are the key elements. The ways to decrease risks and safely protect sensitive information are to develop improved Cybersecurity methods such as multi-factor authentication and the conduct of security audits frequently. Be safe and be vigilant.

Reference:

• https://techcrunch.com/2024/03/28/activision-says-its-investigating-password-stealing-malware-targeting-game-players/
• https://www.bleepingcomputer.com/news/security/activision-enable-2fa-to-secure-accounts-recently-stolen-by-malware/
• https://cyber.vumetric.com/security-news/2024/03/29/activision-enable-2fa-to-secure-accounts-recently-stolen-by-malware/
• https://www.virustotal.com/
• https://otx.alienvault.com/

‍

Introduction

‍

We inhabit an era where digital connectivity, while empowering, has also unleashed a relentless tide of cyber vulnerabilities, where personal privacy is constantly threatened, and crimes like sextortion are the perfect example of the sinister side of our hyperconnected world. Social media platforms, instant messaging apps, and digital content-sharing tools have all grown rapidly, changing how people communicate with one another and making it harder to distinguish between the private and public domains. The rise of sophisticated cybercrimes that use the very tools meant to connect us is the price paid for this unparalleled convenience. Sextortion, a portmanteau of “sex’ and “extortion”, stands out among them as a particularly pernicious kind of internet exploitation. Under the threat of disclosing their private information, photos, or videos, people are forced to engage in sexual behaviours or provide intimate content. Sextortion’s psychological component is what makes it particularly harmful, it feeds on social stigma, shame, and fear, which discourage victims from reporting the crime and feed the cycle of victimisation and silence. This cybercrime targets vulnerable people from all socioeconomic backgrounds and is not limited by age, gender, or location. 

‍

The Economy of Shame: Sextortion as a Cybercrime Industry

‍

A news report from June 03, 2025, reveals a sextortion racket busted in Delhi, where a money trail of over Rs. 5 crore was identified by different teams of the Crime branch. From synthetic financial identities to sextortion and other cyber frauds, a recipe for a sophisticated cybercrime chain was found. To believe this is an aberration is to overlook the reality that it is symptomatic of a much wider and largely uncharted criminal framework. According to the FBI’s 2024 IC3 report, “extortion (including sextortion)” has skyrocketed to 86,415 complaints with losses of $143 million reported in the United States (US) alone. This indicates that coercive image-based threats are no longer an isolated cybercrime but an everyday occurrence. Sextortion is no longer an isolated cybercrime; it has metamorphosed into a systematic, industrialised criminal enterprise. Another news report dated 19th July, 2025, where Delhi Police has detained four people suspected of participating in a sextortion scheme that targeted a resident of the Bhagwanpur Khera neighbourhood of Shahdara. The suspected people were allegedly arrested on a complaint wherein the victim was manipulated and fell prey to a dating site. 

The threat is amplified by the usage of deepfake technology, which allows offenders to create obscene content that looks believable. The approach, which relies on the stigma attached to sexual imagery in conservative societies like India, is that victims frequently give in to requests out of fear of damaging their reputations. The combination of cybercrime and cutting-edge technology highlights the lopsided power that criminals possess, leaving victims defenceless and law enforcement unable to keep up. 

‍

Legal Remedies and the Evolving Battle Against Sextortion

‍

Given the complexity of these crimes, India has recognised sextortion and similar cyber-enabled financial crimes under a number of legal frameworks. A change to recognising cyber-enabled sexual exploitation as an organised criminal business is shown by the introduction of specific provisions like Section 111 in the Bhartiya Nyaya Sanhita (BNS), 2023, which classifies organised cybercrimes including extortion and frauds which fall under its expansive interpretation, as a serious offence. Similarly, Section 318 (2) criminalises cheating with a maximum sentence of three years in prison or a fine, whereas Section 336 (2) makes digital forgery a crime with a maximum sentence with a maximum sentence of two years in prison or a fine. In addition to these regulations, cheating by personation through computer resources is punishable by the Information Technology Act, 2000, specifically Section 66D, which carries a maximum sentence of three years in prison and a maximum fine of Rs. 1 lakh. Due to issues with attribution, cross-border jurisdiction, and the discreet nature of digital evidence, enforcement is still inconsistent even with current statutory restrictions. 

The government and its agencies recognise that laws achieve real impact only when backed by awareness initiatives and accessible, localised mechanisms for redressal. Several Indian states and the Department of Telecommunications launched numerous campaigns to educate the public about and safeguard their mobile communication assets against identity theft, financial fraud, and cyberscams. Initiatives like Cyber Saathi Initiative and Cyber Dost by MHA, with the goal of improving forensic and victim reporting skills.

‍

Conclusion

‍

At CyberPeace, we understand that the best defence against online abuse is prevention. Our goal is to provide people with the information and resources to identify, avoid and report sextortion attempts like CyberPeace Helpline and organise awareness campaigns on safe digital habits. In order to remain updated with the constantly looming danger, our research and policy advocacy also focus on developing more robust legal and technological safeguards. 

To every reader: think before you share, secure your accounts, and never let shame silence you. If you or someone you know becomes a victim, report it immediately, help is available, and justice is possible. Together we can reclaim the internet as a space of trust, not terror. 

‍

References 

‍

• https://www.hindustantimes.com/india-news/delhi-police-busts-sextortion-cyberfraud-rackets-6-held-101748959601825.html 
• https://timesofindia.indiatimes.com/city/delhi/delhi-police-arrests-four-for-sextortion-and-blackmail-in-shahdara/articleshow/122767656.cms ‍
• https://cdn.ncw.gov.in/wp-content/uploads/2025/05/CyberSaheli.pdf