Archive
Format: Research explained
-
Radware expands its AI security offering beyond model-input filtering
The February 2026 product coverage describes vendor capabilities, rather than an independently validated guarantee of agent security.
-
Sysdig’s cloud attack report starts with exposed credentials, not a new AWS login flaw
The reported November 2025 intrusion combined credential exposure, permission escalation and suspected AI assistance.
-
Moltbot security review asks what happens when a personal agent holds all the keys
OX Security’s review concerns the project now called OpenClaw and the permissions users give a persistent assistant.
-
Research into exposed model servers raises access-control and misuse concerns
SentinelOne and Censys examined self-hosted deployments; open model weights alone do not demonstrate that a host was compromised.
-
Reported CISA document uploads highlight the gap between AI access and data approval
January 2026 reporting described security alerts over sensitive contracting files; it did not establish a breach of ChatGPT infrastructure.
-
NSFOCUS’s AI threat matrix puts identity and agent workflows alongside model risks
The vendor announcement describes a governance framework and product direction, not a newly discovered privilege-escalation flaw.
-
Simulated file disclosure is not proof that a chatbot read a server file
Resecurity’s prompt-injection write-up highlights the need to distinguish generated output from demonstrated backend access.
-
ZombieAgent research combined ChatGPT data exfiltration with persistent memory
Radware’s reported demonstration illustrates why recurring prompt-injection variants need system-level containment.
-
NCSC warns against treating prompt injection as another SQL injection problem
The agency recommends secure system design rather than assuming one input-handling technique will eliminate the risk.
-
IDEsaster research connected AI prompts with established editor features
The disclosed collection covered more than 30 findings across coding tools, with product-specific conditions and fixes.
-
Anthropic reported Claude Code misuse in a September 2025 espionage campaign
The company attributed GTG-1002 to a Chinese state-sponsored actor and reported a small number of successful intrusions.
-
AI security is becoming an institutional trust problem, CFR argues
A policy essay connects secure AI deployment with public confidence and the resilience of the technology supply chain.