one letter. that’s all an outside user needed to page through every internal name and email in the workspace.
here’s how Kura caught it 🧵
the list itself is ordinary @owasp hygiene. the point is that no one wrote it for this feature.
Kura derived it from your threat model, recognized broken access control, and attached exactly these controls to exactly this task.
see Kura in action: clover.security/blog/introdu…
what makes the code secure isn’t how much guidance the agent gets.
it’s whether that guidance is about the task in front of it, and grounded in how your system is actually built.
Kura has been running in production across enterprise teams for months.
for every change their agents make, they can see what it decided, where it introduced risk, and how Kura handled it.
works the same whether your agents run on Claude, Codex, or whatever comes next.
see Kura in action: clover.security/blog/introdu…
that threat modeling skill you added to Claude code isn’t securing your code the way you think it is.
so we built something that does. meet Kura 🤝🧵
we tested it.
@owasp, @OpenAI, and @AnthropicAI's skills prevented 32-36% of planted flaws in our controlled research. context scoped to the feature prevented 84%.
the instinct is to give agents more security context: skills, checklists, policy files committed to the repo.
it backfires. the agent compresses all of it and the actionable part gets summarized away.
we tested coding agents on @kubernetesio.
with no security context: 2% of planted flaws fixed.
with feature-tailored context: 74%.
2 → 74. same feature, same prompt, same starting commit. the only thing that changed was the type of security context we gave it.
full research here: clover.security/blog/looks-s…
we ran @OpenAI’s threat modeling skill, @AnthropicAI's and @owasp's on 3 production codebases.
all of them land between 29% and 36% of flaws prevented.
context written for the specific feature being built? 84% of flaws prevented.
we gave coding agents threat models, security skills, @owasp checklists.
17% of planted flaws went to → 33%
then we gave them context written for the exact feature they were building. ↑ 84%.
it’s not about how much context. it’s *which* context you give the agent.
full research here: clover.security/blog/looks-s…
what happens to security teams when AI is doing more of the work?
we’re teaming up with @NotionHQ at their SF office to get into it.
featuring Notion's Jacob Brackett & David Tsao, @vercel's Talha Tariq and our CPO Or Chen.
RSVP: luma.com/of81psme
