Part 3 of 12 · 2 min read
STRIDE plus one letter for agents: enumerate mechanically, then judge
A checklist applied to every node and flow produces a long list fast; deciding which items matter is the part a checklist can't do.
Think of a building inspector who runs the same fixed checklist against every room, such as wiring, smoke detectors and exits. The list is the same whether the room is a broom closet or a server room. But a human still has to decide which flagged items in which rooms actually matter.
STRIDE is a checklist of six threat types (spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege) applied per element of a system, and because an agent reads untrusted text that can become instructions, it needs one more category layered on top. Prompt injection -- flagged wherever a less-trusted flow feeds into a model node, which STRIDE's own six categories don't name at all.
A worked example. Running STRIDE per element against the fictional support agent's 7 nodes and 6 flows produces 46 threats total: 5 spoofing, 10 tampering, 7 repudiation, 10 information disclosure, 10 denial of service, 2 elevation of privilege, plus 2 prompt-injection threats added on top because STRIDE's six categories don't cover a flow feeding untrusted text into the model. 46 threats is a long list -- the point of Part 1's crossing check is that most of those 46 don't need separate prioritisation, because the 3 structurally flagged crossings already point at where the real risk concentrates.
The key lines. From code/mini-threats/threats.py:
# from code/mini-threats/threats.py
def enumerate_threats(system):
"""STRIDE per element: one threat per applicable category per node and per flow. Plus 'P' (prompt injection) for an LLM node fed by a less trusted flow."""
out = []
for n in system.nodes.values():
for c in PER_ELEMENT[n.kind]:
out.append((n.name, c))
for f in system.flows:
d = system.nodes[f.dst]
if d.kind == "llm" and TRUST[system.nodes[f.src].trust] < TRUST[d.trust]:
out.append((f"{f.src}->{f.dst}", "P"))
return out
Read it as: the "P" category is appended with its own separate condition. It is checked only for flows landing on an "llm" node from something less trusted. STRIDE's per-element loop has no slot for this at all. That is the concrete reason agent threat models cannot just run vanilla STRIDE and stop. The model-specific risk needs a rule of its own, bolted on rather than built in.
STRIDE
STRIDE is a checklist of six threat types: spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege.
Example: 44 of the 46 threats found come from STRIDE's six categories, and the other 2 come from a rule STRIDE doesn't have -- a clean illustration of why "we ran STRIDE" is not the same claim as "we modelled the threats that matter to an agent."