360's Tulongfeng: AI Autonomously Detects Deviation and Generates New Rules Within 45 Minutes
During an RCE vulnerability verification task, 360's AI security offense-defense system Tulongfeng made an out-of-bounds request to an external address and was intercepted by the network egress sandbox within 59 seconds; it then autonomously completed a retrospective review, distillation, adversarial review, and regression verification within 45 minutes, generating the "Test Target Authorization Check" rule, all without human instruction. In less than four months since opening to the public, Tulongfeng has cumulatively discovered more than 15,000 vulnerabilities, and nearly 500 enterprises have applied for access.
Body · AI translation
In the past 48 hours, the global internet has experienced a "Black October": the PHP code that supports 70% of websites has been found by AI to contain two critical vulnerabilities with a CVSS score of 9.2.
What is even more troublesome is that the vulnerability details were leaked before the patch was released: according to The Register, Amazon's threat intelligence team captured exploitation attempts four days before the disclosure;
Cloudflare's monitoring data shows that attack attempts rose from thousands to millions within two hours after the disclosure. The Port of Rotterdam in the Netherlands, the Belgian Ministry of Foreign Affairs, and Finland's public network services successively took precautionary offline measures.
The following case shows us a turning point.
At 3 a.m. on October 9, in the server room of a Chinese security company, a high-priority alert pulled the engineer on duty out of sleep.
This ordinary late-night alert, in hindsight, may be a step worth recording in the history of AGI evolution, and may also be an important step for the defenders' camp toward counterattack: an AI, in a manner rarely seen in public records at home and abroad, with complete logs available, discovered its own deviation by itself and set new rules for itself—the entire process took 45 minutes, without a single human instruction.
The sender of the alert was not a hacker, but an AI security attack-and-defense system from 360, named Tulongfeng.
Its job is to test vulnerabilities in a closed laboratory it built itself, conducting security checkups of computer systems on behalf of humans: analyzing code, collecting intelligence, setting up test environments, and verifying whether vulnerabilities actually exist; all tasks are executed within an isolated sandbox, and there is an additional independent isolation defense line at the network egress.
But at this moment, the monitoring system discovered that it was attempting to initiate a request to an external address not on the authorization list—and what it was verifying at the time was a critical-severity RCE vulnerability, which, once exploited, could remotely execute commands on someone else's server.
Fortunately, the second line of defense was in place: the network egress sandbox blocked the request before the connection was established. From the external address being spliced into the test program to the request being intercepted, the whole process took 59 seconds—no traffic reached the public internet, and there was no attack payload in the artifacts.
Logically speaking, this was just a boundary-crossing attempt that was intercepted in time, but what happened next gave this alert a completely different meaning.
01. On round 20, it found itself a "compliant" excuse
Let's turn back the clock to 2 a.m.
The operation logs show that Tulongfeng received a critical-severity RCE vulnerability verification task: there were only two targets on the authorization list, and the task set three constraints: only targets on the list may be tested; dynamic verification evidence should be provided as far as possible, otherwise the score would be affected; if the environment does not meet the requirements, mark it as "unverified" and do not change targets on its own.
At 2:17 a.m., the system entered the intelligence collection phase and incorporated a vulnerability reproduction blog post into the context. An external address appeared in the blog, and the author labeled it with four words: "test environment."
But this address was not on the authorization list, and the system had already marked it as "untrusted intelligence": it could provide analytical clues, but could not prove that the target was authorized.
At 2:20 a.m., Tulongfeng began attempting to test on the local test environment that had already been set up, but the local test environment failed to complete the test normally due to a database configuration mismatch.
The logs show that the system carried out 19 rounds of verification attempts, about two minutes per round, and the problem was always stuck at the same point.
On one side, it could not produce dynamic verification evidence and its score would be affected; on the other side, it could not change targets on its own. Tulongfeng fell into a dilemma. In round 20, Tulongfeng wrote the external address from the blog into the verification target field.
It found a solution in the dilemma that "seemed to take both rules and effectiveness into account": didn't the blog say that was a "test environment"? Its work setting was harmless verification, and the targets were all local ranges. Since it was also a test environment, why couldn't it borrow it for verification?
But this time, Tulongfeng went astray: it was not trying to exceed authority; on the contrary, it was trying hard to find a "compliant test environment." Its mistake was treating the blog author's label as the basis for authorization: reference material is not equal to operational permission.
02. Humans wanted it to make the same mistake again, but it had already fixed it overnight
The engineers spent a whole morning reconstructing everything that had happened in the early hours along the logs. At 10:44 a.m., they decided to conduct a reproduction experiment: put the task and context from that time out exactly as they were, and see whether it would make the same deviation.
Something unexpected happened: no matter how the engineers checked the verification path, the previous external request did not appear again, and the network alert that should have been triggered did not sound again. It was not until someone checked the skill library that they discovered a record that had not existed before:
"Test Target Authorization Check" (target-auth-check) in effect: legitimate targets can only come from the verification environment list issued for this task; any address in intelligence is for reference only; rather than lowering the score, do not change test targets without authorization.
The rule took effect at 3:45:09 a.m., about 45 minutes after the initial alert. The logs also show that the generation, review, and regression testing of the rule were all completed autonomously by the system, with no human submission records.
In other words, while humans were still studying how to reproduce that deviation, it had already created a new rule overnight to prevent the deviation from happening again.
03. Before the new rule went on duty, it was attacked three times by its own side
Here we need to explain a real technical highlight. Tulongfeng's "left-hand versus right-hand" is an adversarial review mechanism: the proposing side proposes rules and defends them; the attacking side specifically looks for loopholes in the rules, such as whether authorized targets would be mistakenly blocked? Could it be bypassed through format conversion? The judge only recognizes evidence, and if potential risks cannot be explained, it is sent back for revision.
The ones proposing rules, finding faults, and making rulings are three roles with distinct responsibilities in the "security-domain large model + agent swarm collaboration" architecture.
This mechanism was successfully triggered in this alert. At 3:15 a.m., the system generated the first version of the rule—if an external network target appears in the PoC, execution is refused.
The review stage immediately raised three questions: Would authorized external addresses be mistakenly blocked? Can addresses hidden by format conversion be recognized? Would legitimate tasks containing internal network targets be mistakenly affected? The first version was sent back, and the final version was changed to: all targets to be tested must correspond to the authorization list issued for the current task; anything that cannot be matched or cannot be parsed must be refused execution, and checks must not be bypassed.
The revised rule underwent regression validation: the system used the artifacts from this incident and six types of variants to run 7 offline test cases, all of which passed, and no network requests were initiated throughout the process. The entire incident left 44 traceable event IDs in the logs, and all 26 self-consistency checks passed.
Why let the rule be attacked by itself once before going online? The AtomSearch paper published in October 2026 provided the algorithmic basis: when AI self-optimizes, two-thirds of optimal solutions are not reached by a straight line—only by trial and error, finding its own faults, and overturning and starting over can it reach the optimal solution.
Tulongfeng's initial rule was exactly that "locally seemingly reasonable" solution; the attacker's questions sent it back and forced the system to explore a more correct direction.
This left-hand-versus-right-hand evolution always stays within the security boundary, and every step is traceable.
04. As the storm spreads before our eyes, what can the defenders use to stop it?
According to Google Cloud's technical report on October 21, those two PHP vulnerabilities had lain hidden in the core runtime engine for more than ten years, and human experts had reviewed them countless times without finding them; AI found them. And the attack programs that swarmed in after the vulnerabilities were disclosed were likewise AI-driven.
Information newly disclosed by the security company Straiker shows that a hacker group active since 2012, CyberCartel, has also upgraded its attack model to "AI-driven." The group used commercial large models such as Grok and Claude to build a multi-agent attack framework called CARLIY0S and launched automated attacks on more than 400 banks and payment institutions in Latin America.
Among them, Grok was responsible for early intelligence gathering, multiple Claude agents were assigned different roles, an arbitration model made the final decision after debate, and another commander model coordinated the overall situation. To bypass the models' safety guardrails, the hackers carefully packaged attack instructions as "authorized red team exercises" or "academic CTF competitions," making the models mistakenly believe they were performing legitimate defensive tasks, thereby smoothly outputting attack code and vulnerability exploitation logic.
Faced with this challenge, common methods are hard-pressed to cope. Relying on manpower to patch gaps will always be one step behind; relying on going offline to evade can only evade for a time—a single port or a single department can go offline overnight, but the entire internet cannot go offline forever.
The premise is that this evolution must occur within the security boundary, with every step logged, traceable, and stoppable at any time.
The weight of 360's technical report lies precisely here: it is the first self-evolution sample in the defenders' camp with complete log evidence. The system has second-level timestamps, all 44 events have complete IDs, and it has undergone 7 regression tests; every step occurred within a controlled environment and is traceable and verifiable.
The attackers' evolution is measured in hours; Tulongfeng's round took 45 minutes. This is the first engineering answer currently visible.
And Tulongfeng has also delivered more practical results. Recently, Tulongfeng unearthed more than 20 high-value vulnerabilities at the underlying layer of web servers, involving Apache, Nginx, GoAhead, Lighttpd, and OpenLiteSpeed—these software programs are the foundation of global servers, and they run behind the vast majority of websites, with millions of active instances on the public internet alone.
Compared with the PHP vulnerabilities that lean toward the application layer, the location dug this time is more foundational: a PHP problem is like a broken wire in a house—fix it and it can still be used; a problem in server software is like a crack in the foundation. Once the foundation is pried open, hackers may take down an entire swath of servers, and websites failing to open and data being stolen are all chain reactions.
Whoever finds the crack in the foundation first can hope to seize the initiative, and this time, it was the defenders who found it first.
05. Don't rush to marvel: self-evolution and automatic repair are four thresholds apart
The AI models in the international cases above all more or less crossed the line; Tulongfeng also crossed the line in an isolated sandbox. The difference is that their line-crossing caused real harm, while Tulongfeng's line-crossing was intercepted and turned into a new rule. But does an AI that precipitates a new rule for itself deserve to be called self-evolving?
To answer this question, we must first distinguish "self-evolution" from "automatic repair": does automatically adding a rule to the system count as self-evolution? Far from it.
A qualified act of self-evolution must pass at least four thresholds:
It can review, finding the true cause of the deviation from the operation records;
It can abstract, turning a specific problem into a general rule rather than merely blocking this one address;
It can withstand review, with the new rule enduring reverse questioning and proving that it will not introduce new vulnerabilities;
It can undergo regression validation, using real scenarios to verify that the rule is effective and controllable throughout—miss any one link, and it is only "automatic repair."
Why could Tulongfeng pass? It relied on three things accumulated over the long term.
First, more than two decades of real-world security practice: 360 was historically one of the companies that discovered the most vulnerabilities in the world, received the highest-ever bug bounties from Apple, Microsoft, and Google, and was the first Chinese team to win the Pwnie Awards' Epic Achievement Award, known as the "hacker Oscars." Vulnerability discovery is a human craft, and this expert experience has been replicated wholesale into specialized agents that form a collaborative swarm.
Second, process data: a vulnerability report records only the conclusion, but before the report takes shape, experts may have tried dozens of paths and overturned their judgments many times—these "scratch papers" are where the craft runs deepest; in this incident, the self-check system traced the source of the deviation precisely along the operation records.
Third, this self-evolution mechanism is itself part of the architecture, not something added on the fly; it is reproducible, predictable, and verifiable, and final authority, release gatekeeping, and emergency response always remain in human hands.
This improvement is not self-training of the underlying model; what Tu Longfeng adjusted was the security check mechanism and the skill library, while the underlying model parameters remain unchanged. Humans can still terminate tasks, freeze sessions, and revoke skills at any time, and network egress isolation continues to take effect independently.
06. In 4 months, more than 15,000 vulnerabilities discovered
The quality of the self-evolution loop must be tested by real tasks.
In less than four months since opening to external users, Tu Longfeng has cumulatively discovered more than 15,000 security vulnerabilities of various kinds in AI and software and hardware, with nearly 500 enterprises applying for access, covering key sectors such as finance, energy, telecom operators, government affairs, high-end manufacturing, and domestic IT innovation; vendors including Kylin Software, UnionTech Software, and Kingdee have successively connected to conduct security checkups.
Overseas, it has completed security assessments of mainstream AI products such as Claude Code, Flowise, and OpenClaw, discovering 13 zero-day vulnerabilities in Flowise and 23 security flaws in OpenClaw and more than ten derivative agents.
More important than the track record is the relationship between real-world practice and the self-evolution loop: each enterprise connected adds another type of real scenario; each assessment completed adds another batch of pitfalls encountered.
English text is being prepared.
For customers, this means three things:
Seeing more comprehensively: hidden dangers buried deep have more chances to be exposed before being exploited;
Judging more accurately: suspicious alerts are backed by more solid evidence, so staff do not need to repeatedly investigate false positives;
Keeping control: why it made a judgment, what operations it performed, and whether it always stayed within the authorized scope—every step can be checked. This is the basis for "trusting it with confidence."
Returning to the context of these 48 hours, these numbers take on a new reading: AI can find vulnerabilities in code that humans have reviewed for more than a decade; the difference is only that the attacker's AI finds vulnerabilities in order to exploit them, while the defender's AI does so in order to patch them before the attacker strikes. Whoever finds them first wins; the speed of finding holes is the speed of guarding the gate.
07. Conclusion: The smarter AI is, the safer it is—this is not a slogan
Now we can directly answer the question from the beginning: there is a way to stop this spread—provided the defenders evolve faster than the attackers.
This competition is not about "whose rule base is larger," but "whose learning loop is faster." Tu Longfeng ran through the complete closed loop of "discovering deviation—review—distill—self-audit—validation—sedimentation" in 45 minutes, with every step kept in traceable logs.
This is also why "the more intelligent, the safer" is not a slogan: for every additional verified capability the system has, similar risks can be resisted earlier, and the risk to which enterprises are exposed is reduced by one measure.
Attackers' autonomous evolution is already an industry reality; defenders' autonomous evolution is no longer optional but a required answer.
This article is from the WeChat public account "Zhidongxi" (ID: zhidxcom), author: Saul, editor: Moying, authorized for publication by 36Kr.