Page 121 - CyberSecurityEssentialsEnglish

TABLE 3-5: THE DIFFERENCE BETWEEN MODES OF AVAILABILITY

Mode What is usually available What should not be assumed

An interface or an API and a Full transparency, or the ability to
Closed service
hosted service. run the same version locally.

The model weights, and possibly The training data, a full description
Open-weight inference code or usage of the build, or unrestricted
examples. licensing.

Components, information, and a
license permitting use, study, That every project using the word
Open source AI modification, and sharing open meets the definition or

according to the published allows training to be reproduced.
definition.

Source: Open Source AI Definition 1.0 [14].

Deep Research and Research Agents
Some products use the name deep research to describe agents that carry out multi-step

searching, visit sources, and assemble a report. This is not a standardized scientific category.
Products may differ in the depth of the search, the type of sources, whether the queries are

shown, the accuracy of the citations, and the ability to return to the original page.
Evaluate a research agent by the coverage and diversity of its sources, the correctness of
every citation, the date of the search, whether it shows what it did not find, whether the

source can be opened, and whether it can abstain when the evidence is weak. A long report
is not evidence of good research.

+ For the advanced reader: multi-agent systems
In multi-agent systems the work may be divided among a researcher, a planner, a critic,
and a coordinator. The division may improve organization, but it does not guarantee
independence of judgment. If the agents rely on the same source or the same

assumption, they may agree on the same error.

Agents and the Limits of Autonomy
An agent is not simply a long-running chatbot. Chapter Five names the components a

production agent is built from and the ladder of terms that separates a workflow from an
agent. It is a system that receives a goal, observes a state, chooses a step, calls a tool, reads
the result, and then repeats the loop. It may use email, a browser, files, databases, a

calendar, or code. The value of an agent therefore rises when a task is multi-step, and the
risk rises because an error can turn from incorrect text into a real action.

120