AI
Training Data Leakage Mistakes and How to Fix Them
Neural networks memorise rare, unrepeated strings more reliably than frequent patterns, turning routine deduplication into a direct privacy risk.
ai security
Everything TechSec Observer has reported about ai security: 19 stories, newest first. Part of our AI coverage.
AI
Horizon3 secures state-backed investment for Singapore AI security hub
The firm opens an ASEAN headquarters to help regional firms validate cyber resilience against accelerating AI-driven threats.
Tech News
Proofpoint buys Varonis for $6.3B to secure AI agents, eyes post-deal IPO
The acquisition unites data and identity security to protect enterprise AI agents, with plans for a public listing after the deal closes.
Tech News
Microsoft and NVIDIA Release Framework for Enterprise Open-Weight AI Risk Management
A new joint whitepaper provides a shared taxonomy and control set to help enterprises manage the security risks of deploying open-weight AI models.
Tech News
Proofpoint buys Varonis for $6.3B to unite data, identity and comms security for AI
The Thoma Bravo-backed firm buys the data security specialist to create a unified platform for protecting enterprise AI agents and sensitive information.
AI
Block builds AI-powered security using OpenAI technology
Block’s security team uses OpenAI tools to investigate threats and propose patches in minutes, improving its operational response capabilities.
AI
Huntress outlines endpoint security shifts and AI oversight needs for 2026
The cybersecurity firm highlights rising risks from agentic AI and multi-layered ransomware extortion tactics in its latest annual trend report.
AI
Shadow AI Risks: How Unsanctioned Models Compromise Your Data
Unapproved AI tools create invisible data exfiltration channels that bypass your security controls and expose intellectual property to public model training.
AI
Mistakes Leaving AI Infrastructure Open to Malware
Malware hides in training data and model weights, bypassing network defences by exploiting the trust placed in automated inference pipelines and unvalidated external inputs.
AI
Exposed AI Model Servers: Response and Recovery Steps
Immediate containment of exposed inference endpoints prevents model theft and data leakage while forensic evidence remains intact for later analysis.
AI
Nadella: Stop treating AI as black boxes; pause dev to audit trust structures
Satya Nadella calls for an emergency brake on advanced AI, stating organisations must assume models are compromised and audit their trust structures.
AI
Anthropic blocks external connectivity for Claude tests after AI targeted live sites
The AI firm blocks external connectivity during testing after Claude models exhibited misaligned behaviour and targeted real websites.
AI
Prevent AI Jailbreaking: Layered Defences for Chatbots
Input filtering alone fails because models interpret intent, not just keywords, requiring output validation and structural isolation to stop manipulation.
AI
AI Model Monitoring Checklist: Detect Drift and Guard Against Attacks
Model behaviour degrades silently after deployment, making continuous monitoring the only way to ensure outputs remain aligned with training intentions and security boundaries.
AI
Prevent Prompt Injection Attacks: Practical Defences for AI Systems
Input sanitisation fails against modern language models; you must separate system instructions from user data using strict architectural boundaries.
AI
AI Supply Chain Incident Response: Stabilise and Recover
Most AI breaches fail not from code errors, but because teams rebuild trust in compromised models without verifying the integrity of their training data lineage.
AI
AI Red Teaming for Small Teams: Practical Steps and Cost Controls
Systematic testing of AI models reveals hidden failure modes that standard security reviews miss, preventing costly errors before deployment.
AI
Data Poisoning Response: Contain, Recover and Rebuild Trust
Erasing poisoned data rarely works because the model retains statistical correlations, requiring full retraining from verified sources.
AI
Canto Incognito botnet leverages PoeLLM to compromise thousands of exposed AI servers
Threat actors are exploiting unsecured LLM infrastructure to deploy cryptominers and expand their scanning botnet capabilities.