TechRisk #106: DeepSeek under $30

Plus, Cloud Threat Report, DeepSeek exposed database with sensitive data publicly, Hackers behind the $70M hack of Singapore crypto firm, and more!

Tech Risk Reading Picks

1. Google’s Threat Horizons Report: Google Cloud Security’s H1 2025 Threat Horizons Report highlights evolving cloud threats, with attackers prioritizing data exfiltration, identity exploitation, and evasion tactics. Ransomware and data theft remain persistent risks, but threat actors are refining their methods, making detection and attribution more difficult. Key threats include over-privileged service accounts, hybrid identity exploitation, database attacks, and Ransomware-as-a-Service (RaaS). Groups like TRIPLESTRENGTH escalate privileges to maximize financial gain, while MFA bypass techniques are increasingly used for extortion. To counter these risks, cloud security strategies must emphasize identity protection and data security, with the report offering intelligence and mitigations to help organizations stay ahead in 2025. [more][more-Google_Threat_Horizons_Report]

2. Google agentic AI defenders: Google is ramping up its AI security efforts to counter threats like prompt injection attacks targeting its Gemini AI system. To tackle these risks, Google’s agentic AI security team has deployed automated red team hacking bots that simulate real-world attacks to identify and mitigate vulnerabilities. These bots use advanced techniques, such as actor-critic models and beam search, to refine and optimize prompt injection attempts, ensuring Gemini’s defenses remain robust. By proactively stress-testing its AI with these hacking simulations, Google aims to strengthen security and protect users from malicious exploits in an evolving threat landscape. [more]

3. Vulnerable Llama: A critical security flaw (CVE-2024-50050) in Meta's Llama large language model (LLM) framework exposes AI systems to remote code execution (RCE) risks due to unsafe deserialization of untrusted data. Found in the Llama Stack’s Python Inference API, the vulnerability stems from using Python’s pickle format, allowing attackers to send malicious objects via the ZeroMQ socket and execute arbitrary code. Meta patched the issue in version 0.0.41 by switching to JSON serialization. This incident follows similar AI framework vulnerabilities, including a recent flaw in TensorFlow’s Keras and an OpenAI ChatGPT crawler bug that could be exploited for DDoS attacks. [more]

4. DeepSeek exposed database with sensitive data publicly: DeepSeek, a Chinese AI startup, accidentally exposed sensitive data by leaving an unprotected ClickHouse database publicly accessible, allowing full control over over a million log entries, API keys, chat histories, and backend details. Discovered by Wiz Research during a routine security check, the breach posed significant risks, potentially allowing attackers to manipulate or extract critical information. DeepSeek quickly secured the database after being notified, but concerns remain over whether unauthorized access occurred. This incident highlights ongoing cybersecurity challenges in AI startups, where rapid development often takes precedence over security, raising privacy concerns, especially given DeepSeek’s prominence and past security issues. [more][more-Wiz]

5. Growing threats powered by GenAI: Generative AI is reshaping cybersecurity, serving as both a defensive tool and a weapon for cybercriminals. While 52% of security experts use AI to combat threats, hackers are leveraging it to launch more frequent, sophisticated, and harder-to-detect attacks. This escalating cyber arms race has exposed a disconnect between CISOs and board members, with the latter underestimating the urgency of AI-driven threats and underfunding security efforts. As AI-powered attacks grow, traditional defences are struggling to keep pace, making education and awareness vital. Strengthening passwords, multi-factor authentication, and employee training remain key in mitigating risks for businesses of all sizes. [more]

6. Deepseek AI model under $30: Jiayi Pan, a PhD candidate at UC Berkeley, and his team claim to have recreated core functions of DeepSeek’s R1-Zero AI model for just $30, highlighting a growing shift in AI research towards cost-effective solutions. While DeepSeek itself disrupted the industry with a model trained for mere millions, Pan’s “TinyZero” pushes the boundaries further, challenging the necessity of billion-dollar AI investments by companies like OpenAI and Google. Their small language model, trained on a number operations game, is open-source and aims to demystify reinforcement learning. If such AI breakthroughs can be achieved on a shoestring budget, it raises pressing questions about big tech’s massive AI expenditures. [more]

7. QRNG enhancement: Palo Alto Networks has developed an open framework for quantum random number generation (QRNG) to enhance cybersecurity against quantum-based threats, collaborating with six quantum technology firms, including Anametric, ID Quantique, and Quantinuum. This initiative aligns with NIST’s recommendations, combining post-quantum cryptography with QRNG to fortify digital security. By adopting open standards, the company aims to simplify the integration of quantum entropy into security infrastructures, ensuring resilience against future quantum threats. The QRNG framework will be accessible via GitHub and integrated into Palo Alto Networks' Next Generation Firewalls, setting a new standard for quantum-ready security solutions. [more]

8. Online travel service’s critical flaw: Cybersecurity researchers have revealed details of a now-patched account takeover vulnerability in a widely used online travel service for hotel and car rentals, which could have allowed attackers to hijack user accounts by exploiting a flaw in the OAuth authentication process. By manipulating a "tr_returnUrl" parameter, threat actors could redirect authentication responses to their own site, gaining unauthorized access to victims' accounts and enabling actions like booking or canceling reservations using airline loyalty points. The attack, which was difficult to detect due to its use of legitimate domains, posed a significant risk to millions of airline users, highlighting the dangers of third-party integrations and API supply chain weaknesses. Security firm Salt Labs emphasized the need for stricter security protocols to prevent such exploits in the future. [more][more-Salt_security_article]

Web3 Cryptospace Spotlight

1. Hackers behind the $70M hack: North Korean hackers are suspected to be behind the $70 million exploit of Singapore-based crypto exchange Phemex, according to blockchain security experts. The breach, which began with $30 million in stolen assets, targeted multiple chains, with attackers swiftly swapping stablecoins for ETH to avoid freezing. Security researchers note the attack resembles previous North Korean-linked exploits, with manual transfers and laundering tactics seen in major crypto heists. At least $44 million flowed through a primary attacker wallet, affecting assets across Ethereum, Solana, Binance Smart Chain, and more. Phemex, which still holds $1.8 billion in assets, is working on a compensation plan for affected users. [more]

2. Attacking browser plugins: 24 Jan - Anti-detect browser, AdsPower, detected a hacker intrusion that spread malicious code, compromising some third-party browser plugins (for example, MetaMask) into phishing versions. According to estimates, hackers managed to steal over $3 million, mostly in Ethereum. This occurred through the theft of seed phrases. The phishing version of the MetaMask extension asked users to re-enter their mnemonic passphrase, after which the funds were stolen. The team swiftly responded by blocking the attack, removing risky plugins, and securing the system. AdsPower has reported the incident to Singapore law enforcement and received a police receipt, with investigations still ongoing. [more][more-AdsPower]