Hi, I'm author. A bit of context on why I built this.
I've spent 15 years in enterprise technology as a Solution Architect. When
our teams
started adopting AI agents with third-party skills, I realized we had the
same blind
trust problem that npm had before npm audit existed — except worse, because
agent
skills can execute shell commands, read environment variables, and make
network
requests by design.
After ClawHavoc hit in January, I saw a dozen scanning tools appear in
weeks. All
heuristic. All pattern matching. The leading one literally says in their
docs: "no
findings does not mean no risk." That bothered me.
So I asked: can we do better than heuristics? The answer is yes — formal
analysis
with soundness guarantees. If the analysis says "no violations," the math
proves
the skill cannot exceed its declared capabilities. Not "we checked and
didn't find
anything" — "we proved it can't."
The key insight: I adapted the Dolev-Yao model (1983, originally for
cryptographic
protocol verification) to model attackers in the agent skill supply chain.
Combined
with abstract interpretation over a capability lattice, SAT-based dependency
resolution, and a trust algebra — you get five provable theorems instead of
five
regex patterns.
Honest about limitations: we miss typosquatting (50% detection — needs name
similarity
module) and dependency confusion (0% — needs registry lookup). These are
v0.2. The
paper documents every gap.
Happy to go deep on any of: the formal model, the benchmark methodology, why
SAT
for dependencies, or the trust score algebra. Ask away.
Hi, I'm Varun — the author. A bit of context on why I built this.
I've spent 15 years in enterprise technology as a Solution Architect. When
our teams
started adopting AI agents with third-party skills, I realized we had the
same blind
trust problem that npm had before npm audit existed — except worse,
because agent
skills can execute shell commands, read environment variables, and make
network
requests by design.
After ClawHavoc hit in January, I saw a dozen scanning tools appear in
weeks. All
heuristic. All pattern matching. The leading one literally says in their
docs: "no
findings does not mean no risk." That bothered me.
So I asked: can we do better than heuristics? The answer is yes — formal
analysis
with soundness guarantees. If the analysis says "no violations," the math
proves
the skill cannot exceed its declared capabilities. Not "we checked and
didn't find
anything" — "we proved it can't."
The key insight: I adapted the Dolev-Yao model (1983, originally for
cryptographic
protocol verification) to model attackers in the agent skill supply chain.
Combined
with abstract interpretation over a capability lattice, SAT-based
dependency
resolution, and a trust algebra — you get five provable theorems instead
of five
regex patterns.
Honest about limitations: we miss typosquatting (50% detection — needs
name similarity
module) and dependency confusion (0% — needs registry lookup). These are
v0.2. The
paper documents every gap.
Happy to go deep on any of: the formal model, the benchmark methodology,
why SAT
for dependencies, or the trust score algebra. Ask away.
Hi, I'm author. A bit of context on why I built this.
Hi, I'm Varun — the author. A bit of context on why I built this.