Archive
Format: Research explained
-
NVIDIA’s analytics example shows why generated-code filters need isolation
A red-team case study traces code-execution risk to the application boundary around model-generated Python.
-
Atlas memory-injection report raised questions about persistent attacker instructions
LayerX’s October 2025 research described a path from a forged request to stored instructions, with current exposure unverified.
-
Claude interpreter research used an allowed API destination as an exfiltration channel
Johann Rehberger’s demonstration questioned whether a domain allowlist alone can prevent data leaving a code sandbox.
-
Cisco’s MCP Scanner offers several ways to inspect tool integrations
The open-source project combines rule-based and model-assisted checks; a scan result is one input to a broader integration review.
-
Brave’s screenshot tests show prompt injection can arrive as pixels
Research on AI browsers explored nearly invisible image instructions alongside conventional page-based attacks.
-
NVIDIA showed how dependency installation can bypass an agent’s blocked-command path
A controlled developer-agent test succeeded through a malicious package after a more direct command attempt was blocked.
-
Poisoning study found 250 documents could trigger a narrow model backdoor
The result covered gibberish-output triggers in experimental models, with important limits on what it proves about frontier systems.
-
A chatbot attack chain can cross several ordinary application boundaries
Trend Micro’s walkthrough connects prompt injection with tool permissions and backend flaws, alongside promotion of its security products.
-
DeepMind introduced CodeMender to generate and validate security fixes
The October 2025 announcement describes early vulnerability-repair research, not a security incident affecting CodeMender.
-
NVIDIA’s red team identifies three recurring LLM application security failures
Unsafe code execution, weak retrieval permissions and active output rendering connect model behavior to real application impact.
-
Microsoft assessed likely AI assistance in an SVG phishing campaign
The observed attack used compromised email accounts and disguised SVG attachments; AI involvement remained an assessment.
-
Cloud AI inventories need to include model access and connected data
Qualys describes its inventory approach in a product article, rather than disclosing a new cloud-model vulnerability.