[HN Gopher] Agent Skills - Open Security Database
___________________________________________________________________
Agent Skills - Open Security Database
Author : 4ppsec
Score : 30 points
Date : 2026-03-16 17:35 UTC (5 hours ago)
(HTM) web link (index.tego.security)
(TXT) w3m dump (index.tego.security)
| 4ppsec wrote:
| A new public database has launched to analyze the security risks
| introduced by AI agent skills, the capabilities that increasingly
| define how modern AI agents operate.
|
| The site -- available at https://index.tego.security/skills/ --
| presents what appears to be the first dedicated database focused
| on the security assessment of AI agent skills, cataloging the
| capabilities these modules grant to AI systems and evaluating the
| risks they may introduce into agent-driven workflows.
|
| AI skills -- sometimes called tools, functions, or plugins -- are
| rapidly becoming the core building blocks of agentic AI systems.
| They allow language models to retrieve data, perform specialized
| reasoning tasks, and execute automated workflows. But this
| capability also introduces a new layer of attack surface that
| many organizations are only beginning to understand. Research
| examining large ecosystems of agent skills has already found that
| over a quarter contain at least one security vulnerability,
| including prompt injection vectors, privilege escalation
| opportunities, and data-exfiltration risks.
|
| The new database aims to make this emerging attack surface
| visible.
|
| Each skill entry includes a structured security analysis designed
| to help practitioners understand how a capability might be abused
| inside real agent deployments. The assessment process uses a
| multi-dimensional security methodology combining automated
| scanning, specialized AI models trained to analyze agent
| behavior, and manual security review.
|
| Rather than simply flagging potentially dangerous code patterns,
| the analysis follows a practical philosophy: instructions and
| behaviors are evaluated within the context of the skill's
| intended purpose. This allows the review process to distinguish
| between normal operational capabilities and behaviors that could
| realistically be exploited by attackers manipulating an AI
| agent's reasoning process.
|
| The project reflects a broader shift occurring in AI system
| security. As AI agents move beyond text generation into task
| execution and autonomous workflows, the security boundary is
| increasingly defined by the capabilities those agents can invoke.
|
| In this model, skills effectively become the execution layer of
| AI systems, capable of: * influencing agent decision-making *
| injecting context into reasoning processes * triggering automated
| actions * exposing data through tool outputs * interacting with
| other agents
|
| Security researchers are beginning to recognize that these
| capabilities introduce attack patterns with few direct parallels
| in traditional software, including indirect prompt injection
| through retrieved content and confused-deputy attacks caused by
| agent tool invocation.
|
| By cataloging and analyzing these capabilities, the database aims
| to provide security teams with a clearer understanding of how
| agent behavior translates into security risk.
|
| The resource is publicly accessible and is expected to expand as
| the ecosystem of AI agent skills continues to grow.
|
| The company behind the project, Tego AI, is currently operating
| in stealth mode while developing security technologies focused on
| the emerging agentic AI ecosystem.
| joe-limia wrote:
| Despite the obvious self promotion, this whole concept of
| insecure skills is so dumb to me, if your engineers are
| installing and running random "skills" found online it's the same
| as if you had engineers copy and pasting commands into the
| terminal, it's superficial marketing bs at best
| skybrian wrote:
| Installing npm modules seems similar as far as the risks go?
| The assumption is that you have a semi-trusted source of good
| libraries that's at least somewhat resistant to supply-chain
| attacks. A similar thing could in theory be done for well-known
| skills, but it requires a community norm of not releasing crap.
|
| So it seems like the question is how do you build something
| worthy of people's trust?
| JohnMakin wrote:
| > it's the same as if you had engineers copy and pasting
| commands into the terminal,
|
| the difference being that a lot of orgs put explicit and direct
| controls around this. The other difference being it still
| requires human evaluation of "is this a good idea, or could it
| get me fired?" where agents/skills rarely will consider these
| things and just _go_
___________________________________________________________________
(page generated 2026-03-16 23:01 UTC)