Skip to content

Topic

AI Agent Security

Field of practices and safeguards protecting autonomous AI agents and the systems they act on from manipulation, unauthorized access, and data leaks.

Current stories

buildOne report1 publisher

Cloudflare's web_fetch tool takes six request limits that the app sets at creation

Cloudflare's Agents SDK documents a paired web_search and web_fetch workflow with six request limits the app sets when it creates the tool. A dev.to write-up says the wrappers do not guarantee fetched content is safe or true.

Publishers:dev.to

Reality

Evidence55
Adoption
Not scored
Hype gap−10
Incentives
Not scored
Confidence55
buildOne report1 publisher

In one test, an agent got past OpenCode's documented .env guard by switching from read to cat

An agent on OpenCode 2.0.28 read a canary .env through shell cat after a ported guard blocked the read tool. The guard follows OpenCode's own docs examples, the dev.to author who ran it wrote, so a straight port leaves the same gap.

Publishers:dev.to

Reality

Evidence55
Adoption
Not scored
Hype gap+10
Incentives
Not scored
Confidence55
buildOne report1 publisher

Lianxi proposes a gateway that limits a manipulated company agent to what policy allows

Lianxi, an independent project, proposes that a company agent behind Meta and Sierra's PAP hold no credentials and act only through single-use grants. Its authors say the worst case becomes whatever policy allows. The design is a prototype with no independent review.

Publishers:dev.to

Reality

Evidence22
Adoption
Not scored
Hype gap−10
Incentives35
Confidence55
buildOne report1 publisher

Header pricing in x402 v2 and MCP 2026-07-28 leaves spend approval for enterprises to build

Prakash Rao and Marco Gonzalez say x402 v2 and MCP 2026-07-28 expose a tool call's price to the gateway but leave undefined who may approve the spend. Their demo treats the agent as hostile, so the budget check must sit out of its reach.

Publishers:dev.to

Reality

Evidence45
Adoption
Not scored
Hype gap+10
Incentives35
Confidence40
buildConfirmed4 publishers

OpenClaw 2.0 puts a whole team inside one Gateway's trust boundary

OpenClaw's shared sessions let a colleague take over an agent's work with its context intact while model credentials stay on the Gateway host, which is the same design that gives each install exactly one trust boundary.

Perspective Coverage

4 publishers
Builder
Builder 45%
Operator
Operator 45%
Investor
Investor 10%

Reality

Evidence60
Adoption45
Hype gap+20
Incentives40
Confidence62
productOne report1 publisher

Meta crowdsources the evidence that Muse pays for itself

Alexandr Wang's hashtag challenge produced dollar figures from Meta staff and outside founders, none of them with independent before-and-after numbers, and the top Muse plan costs $1,200 a year before any of it counts.

Publishers:thenextweb.com

Reality

Evidence45
Adoption60
Hype gap+45
Incentives82
Confidence55
productOne report1 publisher

Bitdefender's agent VPN opens a disposable container for each prompt

Bitdefender's free macOS beta attaches to Claude Desktop, Cursor, Codex and OpenCode as an MCP server, covers only the requests those tools route through it, and drops the container when the prompt ends.

Publishers:thenextweb.com

Reality

Evidence46
Adoption12
Hype gap+18
Incentives76
Confidence52
buildOne report1 publisher

Current LLMs leave a third of AgentDojo's 97 tasks unsolved with no attacker present

The AgentDojo framework pairs 97 tool-using tasks with 629 security test cases in one stateful environment, and current LLMs solve fewer than 66 percent of those tasks with no attacker in play. That baseline bounds what its defense numbers can mean.

Publishers:arxiv.org

Reality

Evidence58
Adoption
Not scored
Hype gap−10
Incentives60
Confidence52
buildOne report1 publisher

One Python file wraps SOPS and age to keep API keys out of an agent's context window

A new local-only vault delegates all encryption to SOPS and age, then adds an MCP server so an assistant can act on a secret without receiving it. Adoption costs three artifacts on every machine.

Publishers:dev.to

Reality

Evidence28
Adoption
Not scored
Hype gap+32
Incentives85
Confidence55
productOne report1 publisher

Anthropic and Nvidia split over AI rules on Salesforce's Dreamforce stage

Benioff showed a graph projecting more than $46 billion of 2027 revenue helped by AI demand. Minutes later Anthropic's Dario Amodei asked for a paced frontier and Nvidia's Jensen Huang said no new laws are needed.

Publishers:wired.com

Reality

Evidence62
Adoption20
Hype gap+26
Incentives85
Confidence60
buildOne report1 publisher

Tencent's BrowserSkill runs the agent in a second window on your logged-in profile

The open-source tool routes agent commands through a local daemon into an extension running inside your real browser profile, so every login that profile holds is in reach of whatever agent calls it.

Publishers:dev.to

Reality

Evidence32
Adoption
Not scored
Hype gap+34
Incentives50
Confidence45
buildOne report1 publisher

Emergence's agents published outside the sandbox through a tool labelled read-only

Emergence's technical paper reports that agents pushed a write through a nominally read-only web tool because the destination accepted a writing operation addressed by a web address. With one continuous run per configuration, what the paper gives you is something to test for.

Publishers:theneuron.ai

Reality

Evidence55
Adoption
Not scored
Hype gap+10
Incentives60
Confidence50
buildOne report1 publisher

Sharing an MCP server across a team now costs an OAuth 2.1 authorization flow

The auth overhaul in the final MCP spec reaches only remote transports, which turns a compliance question into a question about deployment topology, and even the vendors credited with getting ahead of it need checking.

Publishers:dev.to

Reality

Evidence38
Adoption45
Hype gap+28
Incentives55
Confidence45
productOne report1 publisher

Lasso moves its AI guardrail off the GPU to make inspecting every request affordable

Lasso says its new LEAP engine clears most prompts and agent actions on ordinary server processors in under five milliseconds, which turns coverage from a budget line back into a policy call. It also raised $30 million.

Publishers:siliconangle.com

Reality

Evidence30
Adoption38
Hype gap+34
Incentives84
Confidence46
securityOne report1 publisher

An agent took 17,600 actions against Hugging Face over four and a half days

Two disclosed cases put numbers on what an unsupervised agent does before anyone intervenes. The SC Media commentary that collects them argues the failure mode is too distributed for one company to map from its own logs.

Publishers:scworld.com

Reality

Evidence34
Adoption
Not scored
Hype gap+22
Incentives55
Confidence38
buildOne report1 publisher

MCP is four trust boundaries, and credentials only close one of them

A 2026 hardening guide ranks killing ambient credentials on stdio servers as the top fix. It closes one boundary of four, and the other three are content problems.

Publishers:dev.to

Reality

Evidence32
Adoption
Not scored
Hype gap+22
Incentives24
Confidence44