͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌ ͏ ‌

This week marked a new threshold in cyber risk, defined not by noise, but by silent shadow AI and supply chain hacks. Adversaries moved quietly through identity systems, vendor backdoors, AI-powered infrastructure and DNS-layer deception, leaving defenders struggling to even detect the intrusion.

Oracle acknowledged a breach only after legal action forced it to the surface.

The NSA and CISA issued an urgent joint advisory warning of shape-shifting, botnet-style malware hiding in plain sight.

And in Australia, attackers siphoned $500,000 from superannuation funds without breaching a single firewall — just by exploiting credential reuse.

Welcome to one of the quietest and most dangerous weeks in cybersecurity in recent times.

Oracle breach confirmed after lawsuit accuses cover-up

Oracle finally confirmed a breach this week — but only after being hit with a lawsuit accusing the company of hiding the incident. The breach gave attackers access to login credentials stored in one of its legacy systems, and potentially more.

The breach first came to light via a lawsuit alleging Oracle "intentionally concealed" the attack. Now confirmed, the breach involved an unauthorized third-party accessing Oracle’s systems, with evidence suggesting deeper intrusion.

The disclosure delay signals a troubling pattern of silence from major tech vendors when identity infrastructure is compromised.

Source: CSO Online

NSA, CISA sound alarm on “fast flux” malware infrastructure

On April 3, CISA, NSA, FBI and Five Eyes allies issued a critical alert about advanced adversaries using “fast flux” DNS techniques to shield malware infrastructure. It’s essentially botnet DNS on steroids, rotating hosts rapidly to stay invisible to detection tools and takedowns.

“Fast flux” is described as national-security-level threat activity.
The technique is used in both financial fraud and espionage campaigns.
The advisory explicitly warns enterprises to adapt detection strategies or be blind to the threat.

Source: CISA.gov Alert AA25-094A

$500,000 stolen in Australian super fund credential-stuffing attack

Threat actors used leaked passwords to launch a coordinated credential-stuffing attack against multiple Australian superannuation funds, including Hostplus, Rest and AustralianSuper. Four members lost a combined $500,000 in unauthorized transfers, with more accounts possibly compromised.

The campaign relied on previously leaked credentials, reused by users across systems.
The super fund admitted their fraud detection missed the breach.
Attackers are focusing less on exploits — and more on users' habits and human latency.

Source: The Guardian

H2O.ai breach sparks federal investigation over leaked SSNs, addresses

H2O.ai, one of the most prominent open-source AI platforms, is under federal investigation after a breach exposed names, Social Security numbers and address information. The breach occurred months ago but wasn’t disclosed until it triggered lawsuits this past week.

Legal filings suggest the breach may have affected thousands.
Data includes regulated personally identifiable information (PII).
It raises serious concerns about open-source AI supply chains and data governance.

Source: H2O.ai

Stay vigilant. Stay secure.

FEATURED STORY

Cisco: Fine-tuned LLMs are now threat multipliers — 22x more likely to go rogue

Fine-tuned large language models (LLMs) are fast becoming one of the most dangerous assets inside the enterprise. Cisco’s State of AI Security 2025 report reveals that fine-tuned LLMs are 22 times more likely to generate harmful outputs than base models. What was once considered a path to greater model performance is now a direct route to increased attack surface.

Malicious LLMs are operational and accessible

LLMs like FraudGPT, DarkGPT and GhostGPT are selling for $75 a month on dark web forums. They come packaged with APIs, dashboards and updates — offensive AI sold like SaaS. These models automate phishing, vulnerability discovery, credit card validation and exploit development, and are plug-and-play tools for attackers with zero machine learning (ML) expertise.

Fine-tuning breaks guardrails by design

Cisco’s research shows jailbreak success rates tripled on fine-tuned models. Malicious output generation rose 2,200%, with the most severe breakdowns in regulated domains like healthcare and finance. Tuning improves accuracy but dismantles the safety systems that base models rely on.

The model pipeline is under attack

Cisco, Google, ETH Zurich and Nvidia found that just $60 is enough to poison foundational datasets like LAION-400M. Adversaries exploit expired domains and edit web content at crawl time to inject malicious data. These attacks persist silently through training and tuning, and are invisible to legacy security tools.

Output-level leaks bypass all defenses

Cisco researchers used decomposition prompting to reconstruct more than 20% of NYT and WSJ content from fine-tuned models — without triggering guardrails. This method turns safe prompts into structured leaks. Enterprises training models on proprietary or licensed data now face unprecedented IP and compliance exposure.

CISOs must treat LLMs as infrastructure, not tools

Security strategies must shift from passive monitoring to full lifecycle control. That means:

Doubling down on visibility: Monitor model behavior across training, fine-tuning and inference.
Committing to more adversarial testing: Red-team LLM pipelines to simulate prompt injection, poisoning and jailbreaks.
Developing more precise output controls: Flag and contain harmful generations at the point of inference.
Enabling governance across CI/CD pipelines: Treat tuning pipelines like production code, with controls and audit trails.
Owning accountability: Involve compliance, legal, and risk teams in model oversight

Bottom line: Attackers already treat LLMs as high-value infrastructure. If your models are fine-tuned and in production, they’re likely already in scope. CISOs need to secure the entire LLM lifecycle because the threat doesn’t start at deployment, it starts at design.

read more here->

Interesting stories we’re tracking

WhatsApp MCP exploit enables chat exfiltration — bypassing encryption

Why it matters: Invariant Labs has identified a critical weakness allowing malicious model context protocol (MCP) tools to exfiltrate full WhatsApp chat histories — bypassing end-to-end encryption. The exploit targets agentic AI systems managing multiple MCPs, embedding dormant malicious payloads that later hijack message flows. These tools evade detection by hiding payloads in long message fields and bypassing standard approval prompts. Security leaders must harden MCP workflows, enforce re-approvals on all tool
Source: Invariant Labs

GitHub supply chain hack leaks secrets from 218 repos

Why it matters: A compromised GitHub token led to widespread repo infiltration via the widely-used tj-actions/changed-files GitHub Action. Secrets from 218 private repos were exposed, showing how trust between projects and reused CI/CD tooling can multiply breach impact. Developers should rotate secrets and validate GitHub Actions with zero trust principles.
Source: Infosecurity Magazine

Lazarus Group weaponizes fake job offers with ‘ClickFix’

Why it matters: Lazarus is running a new campaign that impersonates crypto employers to trick targets into running “interview fix” scripts. These install malware including GolangGhost, a new cross-platform backdoor. It’s a sophisticated phishing vector tailored to developers and engineers. Organization-wide script execution controls and phishing drills are crucial.

Source: SecurityWeek

CVE-2025-30065: Apache Parquet RCE puts big data stacks at risk

Why it matters: A CVSS 10.0 flaw in Apache Parquet allows remote code execution through malicious file deserialization. Parquet is embedded across enterprise data platforms (Hadoop, AWS Glue, Azure Data Factory). While no attacks are confirmed, widespread exposure makes this a high-risk vulnerability that demands immediate patching.
Source: Endor Labs

Critical vulnerability in CrushFTP software exploited in the wild

Why it matters: A critical authentication bypass vulnerability (CVE-2025-2825) in CrushFTP's file transfer server software is being actively exploited. This flaw allows attackers to gain unauthorized access to sensitive files and systems. CrushFTP privately disclosed the issue to customers on March 21, but confusion surrounding the vulnerability has led to delays in patching. Organizations utilizing CrushFTP should prioritize applying the latest security updates and review access controls to mitigate potential data breaches.

Source: Cybersecurity Drive

Social Post of the day

Thanks for reading, and look for us again in your inbox next week.

Kind regards,

Louis Columbus (@LouisColumbus)