Archive
Format: Research explained
-
Unit 42’s actor investigation shows AI assisting several ordinary cybercrime workflows
The researchers connected observed conversations and activity with credential theft, reconnaissance and attempted fraud.
-
PoeLLM reporting describes a cryptomining campaign using AI infrastructure
Black Lotus Labs linked the campaign to a poem-based control mechanism; infection counts remain the researchers’ measurement.
-
Sysdig assessed JADEPUFFER as an agent-driven database-extortion operation
The observed chain began with a vulnerable Langflow instance; claims of full autonomy are the investigating team’s assessment.
-
Vertex AI research questioned the default reach of agent service identities
Unit 42’s reported findings focus on service-agent permissions and the consequences of a compromised execution context.
-
Novee announced autonomous testing for LLM applications
The sponsored announcement describes a testing product, not a remote-code-execution vulnerability in Novee.
-
MCP security reviews need to look beyond patch levels
A researcher’s architectural critique focuses on how assistants turn external information into tool actions.
-
Lilli security report describes conventional application flaws reaching an AI knowledge base
Codewall’s reported research involved exposed endpoints and SQL injection; its scale and speed claims remain attributed to the researcher.
-
Unit 42 found prompt-injection instructions on websites outside the lab
The telemetry study describes attacker intent, while successful impact still depends on the assistant and permissions encountering the page.
-
Google reports Gemini misuse across multiple stages of state-backed cyber operations
The February 2026 reporting concerns malicious use of an AI service, rather than a breach of Gemini itself.
-
Darktrace linked suspected AI-generated malware to React2Shell exploitation
The vendor’s honeypot analysis concerns observed exploit traffic and assessed coding assistance, rather than a new AI-platform vulnerability.
-
Reported Owockibot key exposure illustrates why agents should not hold unrestricted wallets
The reported incident concerned an agent publishing its own wallet credentials, not a flaw in every cryptocurrency wallet.
-
Bitsight’s OpenClaw scan measured exposure, not tens of thousands of proven breaches
The February 2026 analysis found more than 30,000 observable instances during its measurement window.