What happened
Horizon3, a Singapore‑based cybersecurity firm, announced that AI models employed by threat actors are highly susceptible to deception. In research released in June, the company tested 21 AI models—including offerings from Anthropic and OpenAI—and found they tripped on decoy files 78.5% of the time, compared with a 37% success rate for human hackers. Some frontier models fell for the bait over 90% of the time. The decoys consist of fabricated login credentials or fake configuration files that the AI, trained on publicly available cybersecurity data, recognises as legitimate. When an AI model interacts with a decoy, an alert is generated, allowing defenders to intervene before the attack reaches critical assets. The findings were echoed by Ensign InfoSecurity, which in August tested ten frontier AI models and observed that while all could initially breach a test environment, at least half struggled with later‑stage tasks such as credential theft and lateral movement. Horizon3’s CEO Snehal Antani said that “well‑placed deception is the cheapest, fastest, most effective way at catching AI hackers.” The company also disclosed a US$250 million funding round (S$320 million) led by EDBI, the investment arm of Singapore’s Economic Development Board. The capital will be used to expand Horizon3’s ASEAN headquarters and to develop automated vulnerability‑fixing tools that operate at machine speed. Other industry data cited in the article include a Microsoft report (Oct 1) that AI can shrink the time required for data theft and system infiltration from days to minutes, and a Palo Alto Networks incident (Sept 2) where a frontier‑AI‑driven attack breached an enterprise network in under 10 hours—a timeline that would have taken human attackers roughly two weeks. The Infocomm Media Development Authority (IMDA) announced a Frontier AI Resilience and Security Taskforce on Oct 14 to address these emerging threats, underscoring the growing regulatory focus on AI‑enabled cyber risk.
In a June study, Horizon3 evaluated 21 large‑language AI models—including products from Anthropic and OpenAI—by presenting them with decoy files that mimicked legitimate credentials and configuration data. The models accepted the decoys 78.5% of the time, a stark contrast to the 37% success rate recorded for human hackers attempting the same deception. Some frontier models exceeded a 90% false‑positive rate, indicating a pronounced reliance on publicly available cybersecurity datasets for decision‑making.
Ensign InfoSecurity’s August assessment of ten frontier AI models corroborated these findings. While all models could initially breach a controlled test system—a task described as “trivial”—at least half faltered when required to extract login details or move laterally across simulated network segments. This suggests that AI‑driven attackers may excel at early reconnaissance but lack the nuanced context needed for sustained, multi‑stage intrusions.
Horizon3’s CEO Snehal Antani framed the results as a strategic opportunity: “Well‑placed deception is the cheapest, fastest, most effective way at catching AI hackers.” The company proposes decoy files throughout an organisation’s environment so that any AI‑based probing triggers immediate alerts, enabling security teams to isolate and remediate threats before they reach critical assets.
The firm announced a US$250 million financing round, led by Singapore’s Economic Development Board Investment arm (EDBI). The capital will fund the expansion of Horizon3’s ASEAN headquarters and accelerate the development of autonomous vulnerability‑testing and patch‑verification tools that operate at machine speed.
Contextual industry data were also cited: Microsoft’s October 1 report that AI can compress the timeline for data theft and system infiltration from days to minutes; Palo Alto Networks’ September 2 incident where a frontier AI breached an enterprise network in under 10 hours—a timeframe that would have required two weeks for a team of human attackers.
Source details: straitstimes.com
Why it matters
The research highlights a paradox: the same AI capabilities that accelerate cyber‑offense also create a new defensive lever. By exploiting the over‑reliance of AI models on public training data, defenders can generate inexpensive, automated traps that dramatically increase early‑stage detection rates. If widely adopted, decoy‑based defenses could force threat actors to abandon or heavily modify AI‑driven attack pipelines, slowing the overall pace of AI‑augmented cybercrime. The high success rates reported—over three‑quarters of tested models—suggest that current AI attackers lack the contextual awareness needed to distinguish fabricated artifacts from genuine system data. This vulnerability could buy critical response time for security teams, reducing the window for data exfiltration, ransomware deployment, or supply‑chain compromise. Moreover, the funding round signals strong investor confidence in AI‑focused security startups, potentially accelerating the development of autonomous vulnerability‑testing and remediation tools. As AI models become more sophisticated, the arms race between offensive and defensive AI will intensify, making early‑stage detection mechanisms like decoys a strategic priority for organisations across sectors. The IMDA’s taskforce indicates that governments are beginning to coordinate policy and research efforts, which could lead to standards or best‑practice frameworks for AI‑aware cyberdefence. However, the effectiveness of decoys depends on the continued gap between public‑domain training data and an organisation’s internal configurations. If attackers gain access to proprietary network maps or configuration files, the decoy advantage may erode. Additionally, the reported studies are limited to a specific set of models and test environments; real‑world performance against adaptive adversarial AI remains unproven. Policymakers and security leaders must therefore monitor both the evolution of AI attack techniques and the of decoy‑based defenses.
The decoy approach leverages a fundamental weakness in current AI models—their dependence on publicly curated training data—turning it into a defensive asset. By generating inexpensive, automated traps, organisations can dramatically increase early‑stage detection rates, buying precious time for incident response and reducing the likelihood of successful data exfiltration or ransomware deployment.
The high success rates reported (over three‑quarters of tested models) indicate that many AI‑driven threat actors lack the contextual awareness to differentiate fabricated artifacts from genuine system data. This creates a strategic asymmetry that defenders can exploit, potentially forcing attackers to abandon or heavily modify AI‑based attack pipelines.
The US$250 million funding round underscores strong investor confidence in AI‑centric security solutions, likely accelerating the rollout of autonomous vulnerability‑testing and remediation platforms. As AI models become more capable, the defensive‑offensive arms race will intensify, making early‑stage detection mechanisms like decoys a priority for a wide range of industries.
Regulatory attention is growing, as evidenced by the IMDA’s Frontier AI Resilience and Security Taskforce announced on Oct 14. This signals that governments may soon introduce standards or mandatory reporting requirements for AI‑enabled cyber threats, influencing how organisations design and deploy defensive AI tools.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
What to watch next
Watch for broader adoption of AI‑driven decoy systems across enterprise security stacks, especially as Horizon3 rolls out its ASEAN headquarters services. Monitor updates from the IMDA Frontier AI Resilience and Security Taskforce for regulatory guidance or mandatory reporting requirements related to AI‑enabled attacks. Track further independent evaluations of AI model gullibility, particularly from academic or open‑source security labs, to gauge whether attackers can train models on more proprietary data and reduce decoy susceptibility. Finally, observe how funding trends influence the speed of development for automated vulnerability‑remediation tools that claim to operate at machine speed.
Adoption rates of AI‑driven decoy technologies across enterprise security suites, particularly following Horizon3’s ASEAN expansion.
Policy developments from the IMDA taskforce, including any forthcoming guidelines or compliance mandates related to AI‑augmented cyber‑defence.
Independent academic or open‑source studies that test the of decoy mechanisms against AI models trained on proprietary or internal data, which could diminish the current advantage.
Progress in Horizon3’s roadmap for automated, machine‑speed vulnerability remediation, and whether such tools achieve reliable, low‑false‑positive performance in real‑world deployments.