November 20, 2026:


Escape posted its launch announcement on Nov. 5 and issued a wire release at 9 a.m. ET on Thursday, Nov. 19, for Continuous AI Pentesting, a product that attacks the routes each software change can reach. By the company’s own benchmark, it caught 91% of the vulnerabilities those changes introduced while cutting a run from 78 minutes to 4.9 minutes. Escape has not disclosed how it measured either figure, and PCI Security Standards Council guidance says running an automated tool alone does not satisfy the pentest requirement in PCI DSS, the payment-card standard the launch release cites.
The product starts a run on every merge, after a deploy, or on a daily, weekly or monthly schedule, and it covers web applications, APIs and Model Context Protocol (MCP) servers, according to Escape’s product page. Escape says the security leaders it has spoken with report 10,000 to 20,000 weekly commits and lack the budget to grow testing teams to match. A security lead weighing the product faces three questions the release leaves open: what the benchmark measured, what a run does when a change reaches more than 20 routes, and whether per-release AI runs count as pentest evidence for an auditor.
Teams pick one trigger per profile. The Escape GitHub App starts a run on each push to a watched branch, a customer’s continuous integration system can call Escape’s API after a deploy, or a schedule can start one. Escape’s launch post says only one run executes per application at a time, so merges that land mid-run are covered by the next run, and runs run on Escape’s infrastructure so customer builds don’t wait on them.
A whitebox integration reads the commits made since the last run and reports issues the change introduced, then maps the change to routes it reaches, up to 20 per run. Escape’s product page says a change to a middleware or authentication helper creates risk on routes whose own code did not change, and the mapping is meant to catch those. Escape’s agents then attack those routes on the deployed application, and open issues are replayed against the live app in parallel; an issue closes only after its fix is verified.
The attacking agents come from Cascade Iguazu, which Escape released in late October, according to its consistency post from Nov. 3. Escape’s documentation describes the Cascade engine generally, not Iguazu specifically: an orchestrator plans each engagement and spawns specialist worker agents for reconnaissance, exploitation, validation and reporting, a coverage agent tracks tested surfaces and gaps; and an isolated reporter independently reproduces findings. Escape’s consistent after testing one benchmark app with nine seeded flaws across four runs. Cascade found all nine in every run, while Claude Opus 5.5 found six to eight, an average of 75%, and the authors call the benchmark “deliberately simple.”
Scoping is how a run drops from 78 minutes to 4.9 minutes, and it is one of the oldest trades in software testing. Regression test selection re-runs only part of a test suite after a change, when choosing that part costs less than retesting everything. In a 1997 paper, researchers Gregg Rothermel and Mary Jean Harrold defined a selection as “safe” when it picks every test that can expose faults, under certain conditions. The analogy is imperfect, because Escape’s tests are generated by AI agents rather than drawn from a fixed suite, but the trade is the same: speed bought with coverage. Escape’s materials do not say whether a 20-route selection carries any guarantee of that kind, or what happens when a change reaches more than 20 routes.
Escape says runs scoped to a change finished 16 times faster than a full assessment and caught 91% of the vulnerabilities each change introduced; the 78-minute and 4.9-minute figures come from the same benchmark, and 78 divided by 4.9 is about 15.9, consistent with the 16x. The release does not name the applications tested, the sample size or how vulnerabilities were counted. Escape’s product page describes the 91% only as coming “in our benchmark.” The page links a June 30 Escape benchmark, but that post covers full assessments and contains no diff-scoped runs, no 78-minute or 4.9-minute figure and no 91%.
A 91% catch rate is recall, the share of real flaws a test finds, which means roughly one in 10 change-introduced vulnerabilities went undetected in the benchmark. Recall cannot be read alone, because flagging every item produces perfect recall; precision, the share of reported findings that are real, is the counterweight. The release reports no precision figure. In its June 30 benchmark of full assessments, Escape reported a 0% validated false-positive rate for Cascade across 38 black-box and 56 white-box findings, but Escape validated its own findings, and the post is published by the vendor whose product ranks first. Escape also describes its comparison against Aikido and XBOW, measured by Doyensec, as an external reference, not a controlled test.
Independent critics have named the problems that make such figures hard to read. Ethiack CTO André Baptista argues benchmarks score flags as 1 and everything else as 0, rely on fixed answer keys that can count real findings as false positives, and often trust a single run of an agent that is nondeterministic; he recommends repeated trials and reporting trade-offs among discovery, false positives and severity. Ethiack sells a competing agent, Hackian, and promotes its own benchmark.
The release cites PCI DSS v4.0.1, which requires penetration testing at least annually and after significant application changes, as audit context for the product. Card brands enforce the standard by fining acquiring banks, which pass penalties to merchants and service providers under contract, according to Wikipedia’s PCI DSS entry.
The PCI Security Standards Council’s Penetration Testing Guidance, an information supplement published in September 2017 that says it does not replace any standard, describes penetration testing as largely a manual endeavor that may include scanning, and says “simply running an automated tool does not satisfy the penetration testing requirement.” It also says testers must be organizationally separate from the management of the systems under test, that certifications alone cannot establish qualification, and that PCI DSS does not prescribe what counts as a significant change, leaving that call to each entity.
Escape’s answer is optional: customers can add human-made reports from OSCP-certified testers on the same platform. Neither the release nor the launch post explains how automated per-change runs map to the tester qualification and independence requirements. The council’s guidance references PCI DSS v3.2, so it predates v4.0.1 and agentic testing. Whether an assessor accepts agent-run output as penetration test evidence depends on the assessor and the entity’s documented methodology, which the standard requires the entity to define, document and implement.
Aikido Security introduced Aikido Infinite in February, saying each software change triggers agentic pentesting agents that validate exploitability and retest, and an Aikido survey of 500 security and engineering leaders found 76% deploy significant production changes weekly or faster while 21% validate security on every release. CyCognito said in a release republished July 22 that its Continuous AI Pentesting is now running with select design partners. Intruder launched white-box AI pentests on demand in July for significant releases, with pricing from $4,000, according to its blog.
Escape’s claimed edge is the mapping from a change to routes whose own code did not change, plus AI red teaming for chatbots and MCP servers. Escape’s materials and the coverage cited here include no independent head-to-head comparison of change-scoped tools. Escape does not disclose pricing, and the product is available now directly and through channel partners.
Escape was founded in 2020 by CEO Tristan Kalos and CTO Antoine Carossio, two French engineers who met at UC Berkeley, The Next Web reported; Carossio worked in cybersecurity for the French government before joining Apple as a machine learning researcher. The company announced an $18 million Series A in March, led by Balderton Capital, bringing total funding to about $23 million. Escape both sells the tool and publishes the benchmarks that rank it, which makes the unpublished method behind the 91% figure the deciding gap.
Escape’s AI pentesting documentation says scope is enforced at the network boundary, inside agents and in agent prompts, and that organization administrators can disable all AI pentesting activity through a kill switch. It says customer traffic, payloads, findings and evidence are not used to train third-party models, that zero-retention agreements are in place with commercial model providers where available, and that open-weight models used in sensitive contexts run on Escape-controlled infrastructure. A reporter agent independently reproduces each candidate finding before it is filed.
That page does not address rate limits, guidance on production versus staging targets, or human review of findings. A separate documentation overview lists rate limits among supported guardrails, along with maximum duration, path-level restrictions and strict scope, without giving values. Because the new product’s whitebox integration reads commits from source code, retention and access rules for that code matter, and the release describes them only by saying tests run on Escape’s infrastructure. The documentation describes Escape’s AI pentesting product generally, and the release does not state that each control applies identically to Continuous AI Pentesting.
MCP, the open standard Anthropic announced on Nov. 25, 2024 for connecting AI systems to external tools, was donated in December 2025 to the Agentic AI Foundation, a Linux Foundation-affiliated fund, according to Wikipedia’s MCP entry. The same entry summarizes a May 2026 NSA report saying MCP’s rapid adoption has outpaced its security model.
When a change reaches a chatbot, AI agent or MCP server, Escape’s product page says the run applies AI red teaming: it talks to the chatbot as a user to attack the MCP server behind it, runs multi-turn conversations to hunt for jailbreaks, and checks whether prompt injection leads to unauthorized tool calls, cross-tenant data access or privileged actions. Arkose Labs principal information security analyst Jun Carreon says in the release that Arkose had its MCP server tested as AI-driven attackers would approach it before opening it to customers, and that the results were actionable; he is the only customer quoted.
Prompt injection has no complete fix. The UK’s National Cyber Security Centre said there are no surefire mitigations, and OWASP says retrieval-augmented generation and fine-tuning do not remove the threat. Escape’s own post makes the testing corollary plain: a run that finds nothing new does not prove an application was fully tested.
Escape’s release says, citing Zero Day Clock, that the gap between a vulnerability’s disclosure and its exploitation fell from 32 days in 2022 to hours; that 2022 figure does not appear in the site’s published content reviewed for this article. The Zero Day Clock site charts about 6,070 catalogued vulnerabilities as of Nov. 20.
Other reporting paints different pictures. Hadrian reports a four-hour median for 2024 from Zero Day Clock, a different dataset from its own, and Hadrian’s own Exposure Clock records a 57-day median from disclosure to a listing in the CISA Known Exploited Vulnerabilities catalog. Tom’s Hardware put the mean at one day. Pascal Geenens, Radware’s vice president of cyber threat intelligence, wrote in IT Brief’s opinion piece that mean time to exploit was minus eight hours as of July 23, meaning most tracked vulnerabilities were exploited before disclosure.
The direction is consistent, but none of these datasets measures whether testing each release prevents exploitation. Carossio predicts in the release that by 2027, AI-driven attackers will autonomously find and exploit unknown complex vulnerabilities within minutes.
Escape frames the product as filling gaps between full tests rather than replacing them, and the release keeps a path to human-made reports. Survey data points the same way. Cobalt, which sells pentesting as a service, reported in its June 25 release that the share of 455 surveyed security professionals who would rely entirely on AI automation for testing fell from 29% to 9%, that 78% had experienced critical false negatives from automated tools, and that 47% prefer a hybrid model.
For a security team, the useful reading of this launch is narrower than the pitch. Per-release AI runs can shorten the gap between a merge and a tested route, and Escape’s route mapping targets what the company describes as a blind spot of file-only tools. But the 91% figure is an unpublished recall number with no precision attached, the materials reviewed do not document what a run does when a change reaches more than 20 routes, safeguards and source-code handling are not fully described for this product, and PCI Council guidance still treats an automated tool alone as insufficient. Before buying, ask Escape for the benchmark’s applications, sample size and false-positive rate, what a run does with changes that reach more than 20 routes, and how your assessor should treat the output.
The entity defines, documents and implements its own testing methodology, and its assessor judges whether the evidence meets the standard. The council’s guidance accepts internal or third-party testers if they are qualified and independent of the systems’ management, and it does not treat certifications alone as proof of qualification. Escape’s materials do not say how its automated runs map to those tester requirements.
The PCI Council’s guidance describes a scan as typically automated tools with manual verification of findings, while a penetration test aims to show how vulnerabilities could be exploited. AI pentesting vendors say their agents attempt exploitation and validate findings; Escape says a reporter agent reproduces each candidate finding before it is filed.
Not across a whole application. Escape’s figure covers only the vulnerabilities each change introduced in its benchmark, so it says nothing about flaws already in the codebase, and the benchmark’s applications and sample size are undisclosed.