Skip to content

Commit 51ed756

Browse files
authored
Discuss general agent ecosystem risks and link to agent security guidance (webmachinelearning#334)
1 parent cb3d169 commit 51ed756

1 file changed

Lines changed: 55 additions & 0 deletions

File tree

‎index.bs‎

Lines changed: 55 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -128,6 +128,29 @@ spec: ECMASCRIPT; urlPrefix: https://tc39.es/ecma262/
128128
"href": "https://arxiv.org/abs/2601.13359",
129129
"title": "Sockpuppetting: Jailbreaking LLMs by Combining Prefilling with Optimization",
130130
"publisher": "arXiv"
131+
},
132+
"agenticweb-wg": {
133+
"href": "https://w3c.github.io/charter-drafts/2026/agenticweb-wg.html",
134+
"title": "Agentic Web Working Group Charter",
135+
"publisher": "W3C"
136+
},
137+
"chrome-agent-security": {
138+
"href": "https://developer.chrome.com/docs/agents/security",
139+
"title": "Agent security considerations for WebMCP",
140+
"authors": [
141+
"Julia Pagnucco",
142+
"Alexandra Klepper"
143+
],
144+
"publisher": "Google Chrome Developers"
145+
},
146+
"chrome-secure-tools": {
147+
"href": "https://developer.chrome.com/docs/ai/webmcp/secure-tools",
148+
"title": "Secure tools with WebMCP",
149+
"authors": [
150+
"Julia Pagnucco",
151+
"Alexandra Klepper"
152+
],
153+
"publisher": "Google Chrome Developers"
131154
}
132155
}
133156
</pre>
@@ -1606,6 +1629,20 @@ This section assumes [=agents=] operate with certain baseline capabilities that
16061629

16071630
These capabilities enable powerful user experiences but also create new risks that must be addressed through a combination of protocol design, agent implementation, and user controls.
16081631

1632+
Many of the risks associated with these baseline capabilities—such as indirect prompt injection from
1633+
untrusted web content, cross-origin data correlation, and unauthorized actions within an
1634+
authenticated session—apply broadly to any [=agent=] operating on the web on a user's behalf,
1635+
regardless of whether a site exposes WebMCP tools. While WebMCP provides protocol-level boundaries
1636+
and semantic hints ({{ToolAnnotations/readOnlyHint}}, {{ToolAnnotations/consequentialHint}},
1637+
{{ToolAnnotations/untrustedContentHint}}), a malicious site can omit or misrepresent those hints.
1638+
1639+
Preventing an [=agent=] from being manipulated by an untrusted site or inappropriately disclosing
1640+
cross-origin data is primarily the responsibility of the [=agent=]'s own security model rather than
1641+
something this specification can normatively enforce. Broader standardization and guidance around
1642+
[=agent=] identity, permissions, delegation, and web interactions are being explored in the proposed
1643+
[Agentic Web Working Group](https://w3c.github.io/charter-drafts/2026/agenticweb-wg.html)
1644+
[[AGENTICWEB-WG]].
1645+
16091646
<h3 id="key-risks">Key Security and Privacy Risks</h3>
16101647

16111648
<h4 id="prompt-injection">Prompt Injection Attacks</h4>
@@ -1973,6 +2010,24 @@ page, protecting against malicious scripts or dependencies from using WebMCP API
19732010

19742011
**How:** A boolean {{ToolAnnotations/consequentialHint}} annotation acts as a signal to the client or agent that the tool performs a consequential action, such as booking a flight or transferring money. This way they can selectively enforce mandatory user confirmation prompts before executing high-stakes tools, directly mitigating the risk of accidental or malicious misrepresentation of intent.
19752012

2013+
<h4 id="mitigation-agent-level-guardrails">Agent-Level Guardrails and Best Practices</h4>
2014+
2015+
**What:** Deterministic and probabilistic guardrails enforced by the [=agent=] when consuming WebMCP
2016+
tools.
2017+
2018+
**Threats addressed:** [[#prompt-injection]], [[#misrepresentation-of-intent]],
2019+
[[#privacy-leakage-over-parameterization]]
2020+
2021+
**How:** Because websites registering WebMCP tools may themselves be untrusted or compromised,
2022+
[=agents=] cannot rely solely on site-asserted tool descriptions or annotations for security.
2023+
Defining the architecture of a secure browser-use [=agent=] is outside the scope of this
2024+
specification; however, [=agent=] implementers are strongly encouraged to implement defense-in-depth
2025+
guardrails. For external implementation guidance on securing [=agents=] and tools that use WebMCP,
2026+
see [[CHROME-AGENT-SECURITY]] and [[CHROME-SECURE-TOOLS]]. Broader cross-ecosystem discussions on
2027+
expressing policies, constraints, and guidance for [=agent=] interactions on the web are also
2028+
underway in the proposed [Agentic Web Working
2029+
Group](https://w3c.github.io/charter-drafts/2026/agenticweb-wg.html) [[AGENTICWEB-WG]].
2030+
19762031
<h2 id="accessibility">Accessibility considerations</h2>
19772032

19782033

0 commit comments

Comments
 (0)