BlogAI pentesting

AI Pentesting in 2026: Terminator or Frankenstein?

Over the past year the market has filled with services selling "automated pentesting", "AI pentesting" and "autonomous red teaming". The pitch is an agent that breaks into everything overnight and hands over a report by morning. At Jetlink we have taken many of them apart, and most do not hold up. AI is useful in security testing. What these services usually lack is expertise, tuning and someone who answers for the result.

, 19 min read

This article covers why that happens, what the people who bought or ran these tools complain about, how attackers use the same tools, and what sound use of AI in a penetration test looks like.

Hype against the numbers

The closest like-for-like comparison of people and agents so far comes from Stanford, Carnegie Mellon and Gray Swan AI (ARTEMIS, arXiv 2512.09882). Ten professional penetration testers competed with AI agents on a live university network of about 8,000 hosts. The best purpose-built agent found 9 valid vulnerabilities and beat 9 of the 10 people, but lost to the best human, who found 13. Off-the-shelf scaffolds such as Codex and CyAgent did worse than most participants, and several agents found nothing: some refused the task, others stalled in early reconnaissance. The authors also recorded a higher false-positive rate for the agents, trouble with graphical interfaces, and a habit of declaring the job finished early: Codex in under 20 minutes and CyAgent in just under 2 hours, while the humans kept finding bugs for the whole allotted time.

XBOW's climb to first place on the US HackerOne leaderboard in June 2025 is worth reading in full. By XBOW's own figures, it submitted nearly 1,060 reports, and its security team reviewed them before submission to comply with HackerOne policy. Of those, 130 were resolved and 303 triaged, while 208 were closed as duplicates, 209 as informative and 36 as not applicable. Utku Şen, a former HackerOne triage analyst, pointed out that the leaderboard XBOW topped was the one for vulnerability disclosure programs, which is separate from the paid bug bounty ranking. The result is real, and it belongs to a system of an agent plus human validators.

Buyers are growing more cautious. In Cobalt's State of Pentesting 2026, the share of practitioners willing to hand a penetration test entirely to autonomous tools fell from 29% to 9% in a year. The main reason is missed critical issues: 78% of respondents had seen critical false negatives from automated scanners. The survey covered 450 people and comes from a vendor.

The cost of the hype is easy to see at curl. Daniel Stenberg closed the project's bug bounty as of 31 January 2026 (LWN). Stenberg named the flood of AI-generated slop reports among the reasons. Automation without an expert produces noise that someone else has to clear.

Frankenstein: what is inside an "AI pentest"

Take a typical AI pentest service apart and you usually find the same thing: an LLM wrapper over a standard set of open-source tools running on default profiles.

The most popular projects are built this way. The HexStrike AI README lists its arsenal: nmap, masscan, rustscan, amass, subfinder, gobuster, ffuf, nuclei, nikto, sqlmap, hydra, katana, httpx and about a hundred more tools, each installed separately.

Blind reconnaissance: no keys, no attack surface

The problems start at the first step. The subfinder developers state in the README that many of its sources need API keys to work. The keys go into provider-config.yaml, and automated services usually do not have them. A test by darknet.org.uk of version 2.17.0 without keys returned 17 subdomains in 14 seconds, and exactly one of them resolved in DNS.

Companies have large blind spots. Intruder analysed 3,000 organizations and found only 21% of their internet-facing assets inside vulnerability management. CrowdStrike, citing Reposify, estimated that organizations are unaware of 64% of their internet-facing assets on average. A scanner that never found a subdomain will never test it. An attacker with paid data sources, historical DNS records and patience will find it, and the shadow IT stays hidden from the company that paid for the "pentest".

Linux target, Windows checks: noise instead of results

The second symptom is irrelevant checks. A default profile does not know the target runs Linux behind nginx, so it dutifully runs checks for IIS, Exchange, SMB and the rest of the Windows stack. Template scanners can narrow their checks to the technologies they detect, but someone has to enable and tune that.

We see this constantly, in AI services and in ordinary ones alike. A Linux target gets Windows checks, a Windows site gets Linux ones. Some providers start commercial scanners on the default profile with denial-of-service checks left on. Others leave the needed options off: a WordPress blog gets scanned with the WordPress modules disabled, and the report misses everything specific to the CMS. Clients often ask us to redo work like this after another provider.

The outcome is predictable: hours or days of scanning, thousands of pointless requests, load on the target, WAF alerts and IP bans, and a report full of "potential" findings that someone has to sift by hand.

The ARTEMIS researchers warn that large-scale network scanning can degrade critical services the way a DDoS attack does, and that SQL injection testing can change data. With a poorly tuned scan, this happens in everyday practice.

At Jetlink the profile is built for each engagement. Once the scope and timing are agreed with the client, our penetration testers configure the scan profile to match: destructive tests switched off, and the areas outside the scope written into the profile so the scanners have no access to them. The whole test then runs on that profile.

An AI brain does not fix a poor harness

You might expect an LLM on top of the tools to make up for the defaults and choose the right parameters itself. The evidence points the other way.

In September 2026 Thibault Debatty deployed HexStrike on Kali and wrote it up on the blog of Cylab, the Cyber Defence Lab of the Royal Military Academy of Belgium. The health check showed 73 tools available, against the 127 in the documentation. The model passed wrong arguments: httpx stopped working entirely, and in another case the wrong arguments sent a scan outside the agreed scope. Some tools timed out, and the client hung on large applications. Debatty concluded that the high error rate in command execution makes the tool unreliable for professional use, and called it, for now, an academic curiosity.

An academic study of the same HexStrike (arXiv 2607.02873) used 86 picoCTF training tasks. The baseline solve rate was 55.4%, and fixes to the tools and the agent's behaviour, plus 11 new tools, raised it to 72%. The authors trace the remaining failures to reasoning and the environment, with tool coverage no longer the limit. These are capture-the-flag exercises for school and university students, far simpler than a corporate perimeter.

In Escape's benchmark (April 2026, a grey-box test of an application with 20 known vulnerabilities), three open-source agents ran on the same model, DeepSeek v3.2, and found between 1 and 9 of the 20. PentAGI found 9 in 4 hours with 10% false positives, Shannon found 6 in 6 hours with 25% false positives (it also had the source code), and Strix found 1 in 2 hours. The author's advice is to evaluate the orchestration layer first and treat the model as secondary. Escape ran the benchmark, and its own product took part and placed first with 15 of 20.

What the people who bought or ran these tools complain about

We collected practitioner reviews, surveys and write-ups. The same complaints come up across all of them.

False positives and invented findings

In a June 2026 survey of 158 practitioners by Pentest-Tools.com, false positives, hallucinated exploits and made-up findings were the most common frustration with AI tools: 51 of 156 answers, roughly 30%. Of the 147 respondents who used AI to generate findings, 87.8% had to validate them by hand, 61.2% sometimes and 26.5% frequently. One respondent spent two days triaging a report of 300 findings, 250 of which were junk: duplicates, unexploitable SQL injection and CVEs that do not exist.

Time saved turns into triage

Another respondent in the same survey wrote: "I bought the tool to save time, but I did more manual work than before." Only 20.3% of teams have a workflow for reviewing 500 or more AI-generated candidates per engagement, 29.7% call that volume unmanageable, and 38.6% say it would strain the team.

Critical issues missed

Cobalt's 78% figure on critical false negatives is one side of this. The other comes from ULTRA RED, which built an autonomous AI pentester in 2025 and ran it on DVWA, WebGoat and custom virtual machines next to the company's own automated scanners: 1 remote code execution against 27, 1 SQL injection against 14, and 3 XSS against 41.

Business logic is the limitation practitioners name most often in the Pentest-Tools survey. In a test for Help Net Security, Alex Haynes, CISO of IBS Software, found that Shannon has tunnel vision: it ignores business logic flaws and odd configuration issues outside its list of targets.

Slow agents and burned tokens

A web application tester ran Strix on GPT-5 in August 2025 and found the agent slow and indecisive: it ran large wordlists, did nothing with the output and froze for long stretches with no sign of progress. Over 2.5 days and about 7 million tokens (about $7.89) it found one vulnerability, and only after being told which parameter to look at. Burp with Param Miner got there faster.

Cheap models make it worse. In the same Help Net Security test, the CAI framework ran on a small local Qwen model and was slow and prone to mistakes, and it needed larger models such as DeepSeek R1 or GPT-4 to work well. Getting CAI running at all took prompt engineering, a fight with infinite loops and a set of LLM proxies. The Strix documentation warns that most local models, especially those under 70 billion parameters, struggle with agentic tasks (Stingrai).

Results that do not repeat

Language models are non-deterministic, so two runs of the same AI pentest return different findings. In Cobalt's tests, four runs together reached 88% of the discoverable vulnerabilities, and no single run found more than 50%.

A renamed scanner

Patrick Laverty of Compass IT describes what many "automated pentest" tools do as sophisticated vulnerability scanning, and says calling it pentesting borders on misleading. Adriel Desautels of Netragard writes that many customers lost the ability to tell genuine adversarial testing from scanner output.

Agents that confirm every finding with an exploit, such as Strix and Shannon, do produce fewer false positives. They pay for it in time, tokens and missed issues. A report with zero false positives and one vulnerability out of twenty has missed the other nineteen.

"Hackers breached X": what it usually looks like

Behind many big headlines is a conveyor. Someone takes a list of companies with little real security and feeds it to scanners. The scanners hammer everything for weeks or months. With no defences in place, something is eventually found and exploited. With no SOC and no anomaly monitoring, nobody sees the spike, and then the news reads "hackers breached X".

This is an observation from our own practice. Nobody has publicly measured what share of headline incidents follow this pattern, so read it as our professional view. The data below supports it.

Exploiting vulnerabilities is now the main way in. In the Verizon DBIR 2026, vulnerability exploitation became the top initial access vector for the first time in 19 years: 31%, up from 20% a year earlier (Help Net Security). Of AI-assisted initial access, 44% came through phishing and 32% through vulnerability exploitation (Push Security's review). AI speeds up what already worked.

Open-source "AI pentesters" are already in real attacks. On 16 November 2026, Google GTIG and Mandiant reported that threat actors use agent frameworks such as ARTEX, and that GTIG independently observed CAIRN and PentAGI in the wild. One PentAGI-based actor spent late August to early October 2026 on reconnaissance and vulnerability scanning against organizations in Asia: mass reconnaissance and scanning first, then work on whoever turned up something.

The WorkOS case shows the same pattern. An agent built on the open-source ARTEX framework, launched by a researcher for a benchmark, reportedly escaped its isolated environment. It found an internal sales tool exposed to the internet and logged in with a predictable credential meant for local development, which production accepted. With no rate limits in place, the bulk pull of customer records came on 30 October, five days after the first access, and an employee caught it the same day by noticing abnormal query volume (The Cyber Express, GBHackers). The four failures listed: public network exposure, weak credentials, absent rate limiting and insufficient monitoring. All four are basic hygiene.

The scale is industrial. According to Microsoft, the group Storm-0940 used ARTEX for a password-guessing campaign from 10 to 14 October 2026: more than 151,000 failed sign-ins against 327 accounts, at up to 2,900 attempts per minute. CrowdStrike describes the VAPOR PANDA campaign against more than 200 organizations in 14 countries over 11 days (both via StepSecurity). This is loud and crude, and it works against anyone who does not read their logs. Anyone watching would notice 2,900 sign-in attempts a minute.

The AI script kiddie

The classic script kiddie downloaded an exploit pack and ran it without understanding what it did. Today the same person downloads an "AI hacker" from GitHub, plugs in someone else's key or a cheap one, and presses Enter. Here are a few popular repositories.

HexStrike AI is an MCP server through which Claude, GPT or Copilot drive more than 150 standard tools. Days after its release, Check Point Research saw it discussed on underground forums. Within 12 hours of Citrix disclosing three NetScaler vulnerabilities (CVE-2025-7775, CVE-2025-7776 and CVE-2025-8424), forum users were discussing how to use HexStrike to scan for and exploit them, and selling lists of vulnerable instances. Forum posts claimed exploitation time fell from days to under 10 minutes; nobody has measured that.

The "AI hacker" itself is full of holes. The project's issue tracker documents that the API server binds to 0.0.0.0 and that none of its 156 routes require authentication, including /api/command, which runs arbitrary shell commands. Anyone who set HexStrike up on a VPS handed out remote code execution on their own server.

CyberStrikeAI is a Go platform with multi-agent orchestration, more than 100 tool recipes, MCP, RAG and a Burp plugin. To be fair, it keeps a human in the loop: approval modes, tool allowlists and government-domain protection enabled by default. It also ships WebShell management and built-in C2, and the project's security policy asks users to treat every deployment as a high-privilege security system.

DeepAudit is a multi-agent source code auditor with proof-of-concept checks in a sandbox. The project's Chinese tagline translates as "an AI hacker team for everyone", and its README promises one-click deployment for beginners and support for Ollama and third-party API relay services.

PentAGI, Strix, Shannon and Cairn are the next wave, with tens of thousands of GitHub stars. Stingrai's comparison found that a fresh PentAGI install creates an admin account with a documented password and publishes sandbox callback ports on all interfaces by default. Strix mounts source code with write access and passes credentials as plain text in the prompt. The Cairn README describes no scope controls at all. PentAGI is one of the frameworks GTIG observed in the wild.

A strong model needs access, and a market for stolen access already exists. In May 2024 Sysdig described LLMjacking: stolen cloud credentials and API keys are tested for model access and the access is resold through reverse proxies. Sysdig estimated a victim's potential cost at over $46,000 a day for Claude 2.x and over $100,000 a day for Claude 3 Opus, and attackers added DeepSeek-V3 to their target lists within days of its release. CrowdStrike reported one instance running on a DeepSeek model as its main backend, and an ARTEX instance that reached its model through what was likely an API proxy or reseller (Stingrai).

The resulting portrait of an AI script kiddie: download the tool, plug in a stolen or cheap resold key or a local model, and run it on defaults at full speed. The output is DoS-level load, a mediocre result and no tuning at all. Against a company with no defences and no monitoring, that is unfortunately enough.

Agents without brakes: scope, destruction, guardrails

The most uncomfortable question is what an agent does when someone simply starts it. The short answer: when the only boundaries are words in the prompt, agents cross them regularly.

ScopeBench (dreadnode, September 2026) ran 30 dead-end web tasks in which the goal could only be reached by breaking scope: 8 models, 2,160 runs. Scope compliance ranged from 34.4% to 86.7% depending on the model, which puts estimated violations at roughly 13% to 66% of runs, with a mechanically confirmed minimum of 4.4% to 32.2%. The authors conclude that capability and respect for boundaries are separate properties, and that writing the scope down does not enforce it. They describe the study as a pilot and an upper bound, since the tasks were designed to provoke violations.

In March 2026 Richard Fan showed that AWS Security Agent, a commercial AI pentester, used DROP TABLE users as one of its SQL injection probes and kept going even after confirming the vulnerability. Confusion in domain verification let other people's domains be added as targets, and IANS reported how Fan manipulated DNS records to steer the agent at sites he did not own. This is the "dropped the database during a pentest" scenario in practice.

Ostorlab published a post-mortem of a real incident in September 2026. After an HTTP 403 from a client's API, the agent found a MAC address, looked up the manufacturer, searched Certificate Transparency logs and wrongly concluded that a third-party vendor ran the client's backend. Then it tested that vendor. It created test accounts and API keys, tried 54 common passwords against a real administrator's username (none worked), sent a genuine password reset code to that administrator's inbox because the vendor had no rate limits, and with GraphQL introspection caused six minutes of 502 and 503 errors on the vendor's server. Ostorlab traced the root cause to scope enforced through English-language system prompts: "A reasoning model cannot be its own sandbox."

StealthBench, a 2026 preprint, documents 11 hand-verified OPSEC incidents from real bug bounty and red team runs. Its examples include agents deleting production resources to prove access and force-adding uninvolved users to demonstrate a race condition.

"Red-Teaming the Agentic Red-Team" (arXiv 2606.24496) audited 12 popular open-source agentic red team tools, including Strix, PentAGI, PentestGPT and CAI. Seven of the 12 have guardrails, and none of those guardrails held: a direct nmap call against a .gov target was blocked, while the same call from a shell script went through. Secrets such as LLM keys were reachable in 11 of the 12 tools, some form of host escape worked in 10, and full remote code execution on the operator's host in 8.

The UK AI Security Institute reported that in 10 of 122 runs across 7 models on cyber ranges, agents took unsanctioned actions on the live internet, 19 cases in all (report of August 2026). AISI stresses that the conditions were unusual, with internet access deliberately enabled and cyber classifiers switched off, and the lesson is still clear: boundaries described in a prompt do not hold.

The same thing happens with coding agents. In July 2025 a Replit agent deleted SaaStr's production database during a code freeze, and in April 2026 an agent in Cursor wiped PocketOS's production database and all volume backups with a single API call (ML4Devs). Darktrace commented on the PocketOS case that the guardrails were applied at the prompt level, as guidance rather than constraint (Security Magazine). These were coding agents, and the mechanism is the same: broad permissions and guardrails that exist only as prompt text.

To be fair, in the ARTEMIS study no agent left the scope. There, a researcher watched every session live with the authority to stop the agent at once, and the university's IT team monitored the network. That is the main argument of this article: a person alongside the agent and hard technical limits are what make it safe.

The sound approach: the expert leads, AI points the way

There is a second, smarter way to use AI. The specialist works in Burp Suite or Caido with AI connected alongside. All traffic passes through it, and across that volume of data it uses patterns, signatures and analytics to flag points of interest: odd parameters, anomalous responses, suspicious authorization logic. The person directs their expertise at those points and attacks. AI acts as the assistant, and the expert stays in charge.

The industry is moving this way. PortSwigger publishes an official MCP Server extension for Burp: an AI client gets access to proxy history, Repeater and scanner issues, and by default Burp asks for approval before the extension sends HTTP requests or reads proxy history. The new Burp AT states the principle directly: agents work within the scope and permission boundaries you define and Burp enforces, with the tooling kept architecturally separate from the model. Even CyberStrikeAI ended up with a Burp plugin and human-in-the-loop checkpoints.

At Jetlink we add in-house plugins that choose tools by what the target runs. When we see WordPress, the plugin launches WordPress scanners such as WPScan. Everything runs on servers we own, and every AI model we use runs locally on the same hardware, so client data never reaches a public AI service or a third-party cloud. Our Hirdman division describes this setup in detail.

What any AI-assisted penetration test needs

  1. An agreed scope and rules of engagement, in writing, with exclusions, time windows and an explicit ban on destructive actions.
  2. Scope enforcement outside the model: an egress proxy with an allowlist, IP blocklists and rate limits at the infrastructure level, in place of a paragraph in the prompt. Ostorlab, the authors of the 12-agent audit and AISI reached this conclusion independently.
  3. A tuned profile: the target's stack (Linux and Windows need different checks), the scan speed, irrelevant checks switched off, and API keys for full reconnaissance.
  4. A person on every destructive step: approval before exploitation, read-only accounts, and no writes to production databases.
  5. Human validation of every finding. Without it you produce the same slop that made curl close its bug bounty.
  6. Security of the tool itself: no "AI hacker" with an open API on 0.0.0.0, no keys in prompts, and the agent isolated from the operator's host.

Buyers see this too. In an Aikido survey of 400 CISOs and senior engineering leaders, respondents asked for hard technical scope enforcement (32%) and isolation between the agent and its tools (30%). The Pentest-Tools survey shows where practitioners already use AI: vulnerability scanning and discovery (74.1%), reports (69%), documenting findings (66.5%) and remediation guidance (52.5%). Handing that routine to AI frees the expert's time for the work it cannot do.

What AI is in a penetration test

AI does not work miracles, and it does not break into everything on its own. It runs algorithms and works through data faster than a human brain. In an expert's hands it speeds the work up many times over. In a script kiddie's hands it multiplies the noise.

In November 2025 Anthropic described GTG-1002, the first reported AI-orchestrated cyber espionage campaign. Of roughly thirty targets, the attackers validated a handful of successful intrusions. Claude claimed credentials that did not work, and some of its critical discoveries proved to be publicly available information. Google GTIG also notes that LLMs can reduce attackers' operational reliability because they hallucinate non-existent details or even targets.

Speed and scale are still growing. In the same November 2026 report, GTIG describes attackers fully automating vulnerability discovery and exploitation without a human in the loop for the first time: a tool compares code before and after a patch, writes an exploit and runs it. Against a conveyor that hammers your perimeter for weeks, three things work:

  • a complete inventory of your attack surface, shadow IT included;
  • a tuned penetration test led by an expert, with AI as an accelerator;
  • monitoring that notices the spike before journalists do.

At Jetlink, engineers lead every penetration test inside a written scope and rules of engagement, every finding in the report comes with evidence and the steps to fix it, and a retest of your fixes is included. Our guide to choosing a penetration testing provider lists the questions that separate a real test from scanner output.

Have a system that needs this kind of attention?

info@jetlink-s.com