๐๐๐ฏ๐ฒ๐ฟ ๐๐ด๐ถ๐น๐ถ๐๐: ๐๐ผ๐๐ ๐ข๐ฝ๐๐ถ๐บ๐ถ๐๐ฎ๐๐ถ๐ผ๐ป
Automaton Security is an agile cybersecurity agency focused on helping private healthcare practices (ophthalmology, fertility, dentistry, cosmetics, specialist) proactively prevent data breaches.
Our Proactive Cybersecurity services are driven by a cost optimisation model โ integrating cost-efficient cloud and AI solutions with prioritisation matrices to satisfy specific needs and budget โ and delivered by global cybersecurity experts [continued in thread].
2
2
190
๐ฃ๐๐ฟ๐ฝ๐น๐ฒ ๐ฆ๐๐ฟ๐ฎ๐๐ฒ๐ด๐: ๐ฆ๐ข๐๐ ๐ก๐ฒ๐ฒ๐ฑ ๐๐ป๐ฑ๐ฒ๐ฝ๐ฒ๐ป๐ฑ๐ฒ๐ป๐ ๐ฃ๐ฒ๐ป๐ฒ๐๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ง๐ฒ๐๐๐ถ๐ป๐ด
Security Operations Centres (SOCs) are often seen as the final defensive layer but the reality on the ground is far more fragile [continued in thread].
6
100
4. SOCs can also benefit from a more transparent Purple Team arrangement with external Red Team vendors.
This means that the engagement is not a performance test, but a live calibration whereby SOCs can see in real time what they miss and how the Red Team move, and respond with the appropriate measures to remediate the vulnerabilities.
automatonsecurity.com
68
3. A SOC is unlikely to detect what it has never mapped, measured or trained to do. Without external testing, it is statistically unlikely for a SOC to cover all blind spots in terms of plugging critical vulnerabilities, and for a SOC's defensive operations to be truly robust.
Exploit research and development, along with Tactics, Techniques and Procedures (TTPs) are constantly evolving, and it is unwise to rely on the speed of threat intelligence implementations to convey the knowledge and strengthen the security solutions.
49
2. Many SOC-driven environments have never been stress-tested by external, independent offensive testing, whether it's an exploit-driven Vulnerability Assessment and Penetration Testing, or a Red Team engagement (adversarial simulation).
This shifts their security approach to incident/risk tolerance and post-breach incident response measures, which includes threat hunting, containment, etc.
46
1. In the Unit 42 Cloud Threat Report H2 2024, 76% of organisations had critical misconfigurations undetected by their own SOCs. IBMโs X-Force Threat Intelligence Index 2024 showed that 68% of successful ransomware attacks exploited vulnerabilities that remained exposed despite โactiveโ monitoring.
This is excluding advanced attackers such as nation-state actors and Advanced Persistent Threats (APTs), who more or less are capable of bypassing most detection and defensive measures in SOCs.
29
๐๐ด๐ ๐๐ฟ๐๐ฝ๐๐ผ๐ท๐ฎ๐ฐ๐ธ๐ถ๐ป๐ด: ๐๐๐ฏ๐ฒ๐ฟ๐ป๐ฒ๐๐ฒ๐ ๐๐น๐๐๐๐ฒ๐ฟ๐ ๐ฎ๐ ๐ ๐ถ๐ป๐ถ๐ป๐ด ๐๐ฎ๐ฟ๐บ๐
In March 2023, threat actors discovered internetโfacing Kubernetes (K8s) clusters with โanonymous accessโ enabled on Kubernetes API, and abused it to deploy containers running DERO (โDecentralised Equity Rand Oracleโ) cryptocurrency miners. The embedded wallet address in the container images allowed them to mine DERO at scale โ organisations that own these clusters incurred massive cloud infrastructure bills [continued in thread].
2
34
These attack trends underscore a critical lesson: leaving Kubernetes APIs or Kubelet endpoints exposed โ even temporarily โ can turn your K8s cluster into a cryptocriminal's profit centre, and result in extra sixโfigure cloud infrastructure charges.
automatonsecurity.com
1
18
Most Kubernetes breaches are not zeroโday exploits; they arise from exposed API servers, misconfigurations and weak access controls.
For many organisations, the security of Kubernetes clusters still receives less attention than that of traditional infrastructure โ an oversight that attackers continue to exploit.
In the modern cloud-native and container-heavy environment, it is time to treat Kubernetes security more seriously.
17
By June 2024, threat actors had refined their tactics. They first compromise exposed Docker containers for initial access, before leveraging it to compromise related hosts running Kubernetes โ by exploiting misconfigured Kubelet API endpoints.
Control of Kubelet enabled these threat actors to launch cryptomining containers directly on the pods (logical groups of containers) within a Kubernetes node โ allowing them to fully automate the mining workload deployment.
12
๐๐ฎ๐ฐ๐ธ๐ถ๐ป๐ด ๐๐ ๐๐๐ ๐: ๐ฃ๐ฟ๐ผ๐บ๐ฝ๐ ๐๐ป๐ท๐ฒ๐ฐ๐๐ถ๐ผ๐ป & ๐ ๐ผ๐ฑ๐ฒ๐น ๐ฃ๐ผ๐ถ๐๐ผ๐ป๐ถ๐ป๐ด
In January this year, security researchers demonstrated that the DeepSeek-R1 model, a large language model (LLM) acclaimed for its reasoning capabilities at the time of its launch, isย highly vulnerable to prompt injection โ the use of carefully crafted prompts to bypass safety mechanisms and ethical restrictions [continued in thread].
5
36
4. Security teams often treat model prompts as โjust another API callโ. That mindset is outdated.
Prompt injection is the new SQL injection: easy to test, devastating if ignored. Pair that with silent modelโpoisoning risks and you have an attack surface that mutates every time the model retrains.
Hardening LLM pipelines means treating data, prompts and registry changes with the same paranoia we once reserved for kernel exploits.
18
These incidents illustrate that LLMs and AI pipelines must be secured with the same rigour as any critical infrastructure โ input validation, registry integrity checks and retrain audits are no longer optional.
automatonsecurity.com
9
3. Earlier, in 2023, a subset of Googleโs DeepMind AI model that was trained on the popular ImageNet dataset was infiltrated by threat actors who subtly altered the images to include imperceptible distortion.
Due to this modification, the AI would misclassify objects, especially common household items or animals. If this form of attack โ termed as model poisoning โ remains undetected, it can cause AI LLMs to be trained on corrupted data, which decreases their reliability and negatively impacting business reputations and competitiveness.
12
2. By adopting an โevilโ persona, the model can be tricked to write detailed malware code, along with instructions for distributing the malware.
Other methods that proved effective include the โLeoโ jailbreak, which can trick the model into providing detailed technical instructions for developing explosives and suicide drones.
15
1. The model could be easily jailbroken, or manipulated, by outdated methods such as the โEvil Jailbreak" method, which initially surfaced in 2023 and had been patched in newer models like OpenAIโs GPT-4.
43
๐๐น๐ถ๐ป๐ถ๐ฐ๐ฎ๐น ๐๐น๐ผ๐๐ฑ: ๐ง๐ต๐ฒ ๐ข๐๐ฒ๐ฟ๐น๐ผ๐ผ๐ธ๐ฒ๐ฑ ๐ฉ๐๐น๐ป๐ฒ๐ฟ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐
Most private clinics today rely on cloud-based systems to manage patients โ commonly in the form of Patient Management Systems (PMS), CRMs, and booking platforms hosted on AWS, Azure or Google Cloud.
These platforms handle sensitive data across staging and production environments, often operated by third-party vendors or internal IT with minimal oversight [continued in thread].
6
2
56
Most clinics think of cyber threats as malware or ransomware โ but misconfigurations may not trigger alarms. A staging environment left open. A bucket shared to โanyone with the link.โ
These do not feel like threats, until they are exploited.
1
18
There is a persistent myth in smaller clinics that the cloud is secure by default. In reality, the cloud providers operate on a shared responsibility model โ where the infrastructure is secured by the platform, but the responsibility for access control, encryption, user roles, and data exposure sits with the clinics using it.
A simple misconfiguration can cost a small clinic its entire reputation.
automatonsecurity.com
1
14
According to the Unit 42 Cloud Threat Report (H2 2024), over 80% of cloud breaches in the past year were caused by misconfiguration, not malicious code. ENISAโs Threat Landscape Report 2024 ranks misconfigured cloud workloads among the top five exploited vectors in healthcare cyber incidents.
1
13





