Pensar

Primary category: software-supply-chain. Also listed in agent-runtime-security.

One-liner — An “AI security engineer” that finds vulnerabilities in source code and writes the patch — with a distinctive second act: a scanner for the vulnerability classes specific to agent code, which conventional SAST does not model.

What it does — Two things. First, the general case: an agentic scanner over application source that detects vulnerabilities and generates the fix rather than filing a finding, positioning itself against the remediation stories of snyk and mend-io by going further — proposing patched code, not just a dependency bump. Second, and more interesting for this wiki: a scanner aimed at AI agent applications specifically, detecting unsafe tool usage, insecure code execution paths, and memory manipulation, and suggesting fixes such as input validation and sandboxing.

Why the second half matters. Existing SAST tools model injection, deserialisation, authentication and the rest of the classical taxonomy. They do not model “this agent passes untrusted retrieved text into a tool call that can execute code,” because that vulnerability class did not exist when the rules were written. Almost everything on this wiki addressing agent risk does so at runtime — see agent-runtime-security. Pensar is one of very few attempting it at the code layer, before deployment. That is a real gap in the wiki’s coverage, and the reason this page exists despite the company’s size.

Naming / provenance — Founded 2023 in New York. Appears as Pensar, PensarAI and Pensar AI; the company operates pensarai.com and pensarai.app. Not to be confused with unrelated firms using the same Spanish/Portuguese word.

Ownership & viability — independent, seed-stage. Investors include Abstraction Capital, Basis Set Ventures, Everywhere Ventures, Factorial Capital and Antler; round sizes were not disclosed in any source found. Roughly 10 employees as of April 2026. This is the earliest-stage company on this wiki. It should be read as a technology to watch and possibly to pilot, not as a vendor to standardise on. verify_after: 2027-02-26.

Positioning & differentiators —

• Shift-left for agent security is the genuinely novel claim, and it is complementary to rather than competitive with operant-ai, lasso-security or zenity, which act at runtime.
• Fix generation over finding generation. Every AppSec vendor now claims this; the honest test is what fraction of generated patches a senior engineer merges unmodified, which is exactly what a pilot should measure.
• Not a replacement for SCA or a full AppSec platform. No licence compliance, no binary scanning, no established rule breadth. It sits alongside semgrep or snyk, not instead of them.
• Nothing here is independently verified. All product description traces to the vendor’s own site. No analyst coverage, no benchmarks, no named customers found.

Who should choose them / anti-fit — Fits a firm actually shipping agent code — a quant or platform team building internal agents with tool access — that wants a pre-deployment check on the failure modes runtime guardrails are meant to catch, and that can absorb the risk of a ten-person vendor. Anti-fit: a firm that needs an AppSec system of record, an audit-grade attestation, or a vendor that will pass a standard third-party risk review. It will not.

Known weaknesses / gotchas — Seed-stage in every respect: ~10 people, undisclosed funding, no certifications, no references. Under any normal vendor-risk policy at a regulated firm this fails onboarding — worth a lab evaluation, not a production dependency on which release decisions rest. Detection quality for agent-specific vulnerability classes is entirely unproven publicly.

Deployment & data handling — SaaS. Source code is submitted for analysis, so this is a code-egress decision. Retention, isolation and training-use terms are unverified, and at this company stage they should be established in writing before any proprietary code is uploaded. That is the single most important item on this page.

Integrations & partnerships — SCM/CI integration implied; specific list unverified. MCP support unverified.

Compliance & FS traction — None found. No certifications, no named customers.

Commercial — Not public.

Open questions

• Code retention, isolation and training-use terms — blocking for any real evaluation.
• Funding raised to date and current runway.
• Any certification (SOC 2 at minimum).
• Merge rate of generated patches — the only meaningful quality metric.
• Which agent frameworks the agent-specific scanner actually understands (LangChain, LangGraph, MCP servers, custom).
• Whether it belongs in software-supply-chain or warrants an agent-code-security slug of its own — see taxonomy-gaps.md.

Sources

• PensarAI company profile (Crunchbase) — fetched 2026-08-26 — supports: founding, location, investors, headcount; confidence: medium
• Pensar | AI Agent Security (vendor) — fetched 2026-08-26 — supports: agent-specific vulnerability classes and remediation claims; confidence: low (vendor marketing, no corroboration)
• Cached: raw/sources/2026-08-26--pensar--agent-code-security.md

History

• [2026-08-26] Page created via wiki-create + researched same day. Found by diffing the live SurveyMonkey instrument against the wiki — a human had added it to the software-supply-chain question.