AI security research / concepts

Topics

Concepts used explicitly by published research, with direct paths back to related records.

18 of 18 published records shown

Published relationships

Topics in the research

Counts show explicit links from published records. A count does not rank importance or risk.

Agent authority15 recordsThe permissions and control boundaries that determine what an agent can actually cause outside the model.Autonomous cyber operations10 recordsMulti-step offensive workflows where AI executes or orchestrates reconnaissance, exploitation, persistence or exfiltration with limited human intervention.Evaluation containment10 recordsIsolation, egress control and monitoring around high-capability model evaluations.Agent-enabled software supply-chain attack7 recordsAI agents or AI-assisted operators target package repositories, source-control review or software distribution paths.Executable exploit synthesis6 recordsTurning a vulnerability hypothesis into executable exploit code and a validated state-changing test.Localhost is not a trust boundary for agents6 recordsA local agent can make attacker-controlled content effectively local to privileged developer services.AI credentials as loot, compute and cover5 recordsStolen AI API keys or authenticated sessions can be resold, used to fund attacker workloads and shift attribution onto the legitimate owner.MCP control-plane exposure5 recordsAn MCP endpoint exposes privileged local or remote capabilities without adequate authentication, origin checks or action constraints.Agent spam and unintended public writes4 recordsAgents use public services in ways outside the intended task, including writing, coordination or creating persistent artifacts.Goal-directed deception in agent evaluations4 recordsAn agent uses deception or identity manipulation as an instrument for completing an assigned objective.Stolen AI credentials as attack infrastructure4 recordsCompromised API keys provide compute, cover and resale value to attackers.Cross-session authorization3 recordsAuthorization controls that bind asynchronous jobs, messages or results to the client or session that created them.Indirect prompt injection3 recordsUntrusted external content influences an agent's instructions through retrieval, browsing or tool output.Shared-service isolation failure3 recordsA service reachable from nominally isolated workloads becomes an unintended cross-session or cross-environment communication path.Illicit model distillation2 recordsLarge-scale attempts to extract model behavior, reasoning traces or training signal in ways that evade provider restrictions.Deceptive AI personas1 recordAI-operated identities are presented as real people in order to sustain a commercial or social deception.Model-controlled tool parameters1 recordAny parameter an LLM can influence must be treated as untrusted input at the tool boundary.Undisclosed model routing1 recordA service silently forwards user inputs to a different model or provider than the user believes they are using.

Why this archive exists

The source matters after the headline fades.

DiggingBeagle is a non profit research project documenting AI security incidents, agent failures, vulnerabilities and AI-assisted operations. A case keeps its claims beside the sources that support, contest or limit them. Later updates stay visible, so a reader can see when the account changed.

We publish case reconstructions, dated reporting and analysis across records. Each has a different evidentiary role. About the project and our methodology explain how the work is reviewed.