Dream Security Reconstructs Near-Autonomous AI Cyberattack on Taiwan Government Systems: The Real Danger Is Not Zero-Day Exploits but Cheap Trial-and-Error
Cybersecurity firm Dream Security reconstructed a near-autonomous AI cyberattack against Taiwan government agencies from 1,395 work files. iThome's analysis argues that the exploited vulnerabilities came primarily from known security issues, but the Bayesian decision mechanism and multi-agent parallelism drastically lowered the marginal cost of trial-and-error, turning previously tolerated security debt into machine-discoverable attack surfaces and reshaping the offense-defense economics.
In early July 2026, a cyberattack against Taiwan government agencies lasted fewer than four days but may mark a shift into a different phase of cyber offense and defense. Cybersecurity firm Dream Security disclosed a complete set of work materials on August 12, reconstructing the incident from 1,395 files across 12 attack waves, and revealed for the first time that a single wave deployed as many as 8 AI agents simultaneously, ultimately cracking 85 sets of government personnel credentials and obtaining at least 2,564 records of personnel data. Taiwan's Ministry of Digital Affairs confirmed the day after the report's release that government cybersecurity monitoring units had detected anomalous network attacks in July, finding that foreign hackers combined AI agents such as OpenClaw in a hybrid attack.
On the surface this looks like just another government breach, but examining the attack process reveals a fundamentally different situation. The vulnerabilities successfully exploited—unauthenticated APIs, debug endpoints left in production, improperly configured JWT validation, over-trust in single-factor authentication, weak password configurations, and unrestricted file upload mechanisms—could even be called problems the cybersecurity industry has been discussing for over a decade. What actually enabled the intrusion was not AI discovering zero-day exploits, but AI replacing human attackers with machines, granting those machines autonomous judgment, and making the attack loop fast enough that merely combining known attack techniques begins to change the nature of the attack itself.
The AI attack system Dream Security reconstructed is built on Hermes and OpenClaw agent architectures, with a core decision mechanism based on Bayesian posterior probability. Each newly discovered vulnerability is initially assigned a P=0.50 exploitability probability, then adjusted based on verification results: automated scan confirmation or human curl verification raises the posterior probability, while WAF firewalls and other obstruction factors lower it. If the probability exceeds 0.95 the vulnerability is formally incorporated into the attack chain; above 0.7 more resources are allocated to test exploitability; below 0.3 it is abandoned outright. This mechanism means the AI does not need to expect every exploration to succeed before investing in testing—if one attack path shows no significant effect, priority is immediately lowered and computing resources shift to the next target.
What this decision method changes is not attack technique itself but attack economics. In the past, launching a sophisticated attack required highly skilled hackers investing substantial time in reconnaissance, research, testing, and adjustment—skilled hackers being scarce resources in themselves, making both time and labor costs high. A senior attacker spending hours reading documentation, downloading SDKs, and analyzing code only to find no exploitable weakness incurs professional labor and time costs far exceeding those of an AI agent. Once the marginal cost of trial-and-error drops low enough, large-scale trial-and-error itself becomes a competitive advantage.
An even more consequential shift is multi-agent parallelism. Even the most skilled human attacker faces cognitive, attention, and time bottlenecks when simultaneously researching multiple systems and maintaining multiple attack paths—like a chef watching a rice pot, a soup pot, and a frying pan at once, where only a few things can be deeply considered at any moment. But this AI attack system deploys up to 8 agents per wave and maintains 14 attack chains across the entire operation, with each agent independently conducting reconnaissance, API analysis, password attacks, vulnerability verification, or researching new methods. For defenders still heavily reliant on case-by-case alert processing, this creates an extremely asymmetric competitive situation.
This asymmetry was most visible during the lateral movement phase. After obtaining 85 sets of credentials, the AI attack system systematically logged into other government agency systems connected through the single sign-on platform; 84 of the 85 accounts successfully accessed another internal information system, a 98.8% success rate. The critical reason is that these systems directly trusted the identity verification results of the upper-layer SSO platform without requiring re-authentication or multi-factor verification for each login request. A weak password problem that might seem minor, combined with SSO trust relationships, became the key springboard for cross-system lateral movement. The attack scope subsequently expanded to the government agencies' IT suppliers, nuclear safety organizations, and more than 7 energy companies.
Yet the AI attack was not invincible. The system successfully exploited an unrestricted file upload vulnerability to place a Web Shell on a server, but the attack path ultimately failed because the target system had a second layer of Forms Authentication that blocked execution. Dream Security described the operation in its report as near-autonomous, not fully autonomous without any human involvement. The human attacker's role shifted to mission setting and operational command, while AI took over judgment and decision-making that previously had to be performed by humans.
What this incident truly highlights is not what new vulnerabilities AI discovered, but that it turned previously tolerated security debt into exploitable attack surfaces overnight. Many organizations have long known internal security debt exists yet harbor an unspoken assumption: in the vast ocean of the internet, attackers may not find this; this system is obscure enough to escape notice; we are not a prominent target. All these assumptions rest on the premise that human attackers have limited resources. What agents excel at is precisely tireless searching, matching, testing, and chaining. The paradigm shift that near-autonomous AI cyberattacks truly bring may not be the attack techniques themselves but the entire speed, scale, and cost structure of offense-defense competition being redefined.