AI guardrails built by frontier providers weaken defender advantage, Cisco Talos finds

KUALA LUMPUR, 28 AUGUST 2026 —

AI Guardrails May Backfire Against Defenders, Security Experts Warn

Bianco, writing in the Threat Source newsletter, highlights what he terms the "safety penalty" — a self-inflicted erosion of defenders' inherent advantages. While attackers must evade detection at every stage of an intrusion, defenders only need to spot the threat once to neutralize it. However, AI systems that block or delay investigative actions — even with good intentions — can provide attackers with critical breathing room.

"Agentic security operations center (SOC) processes that refuse to execute or flag actions for human review may halt investigations entirely," Bianco writes. "This gives adversaries time to complete their objectives before defenders regain control." He stresses that the location and flexibility of guardrails matter more than their strictness. Organizations should retain operational sovereignty by customizing AI safety controls within their own environments, rather than accepting rigid, externally enforced policies.

---

Malaysia’s Cybersecurity Stakeholders Urged to Reclaim Control Over AI Tools

The warning comes as Malaysia accelerates its digital transformation, with government agencies, financial institutions, and critical infrastructure operators increasingly adopting AI-driven security tools. While AI promises faster threat detection and automated response, experts caution that over-reliance on vendor-controlled guardrails could weaken local defenses.

"Malaysian organizations must ensure their AI models are aligned with domestic threat landscapes and regulatory requirements," said a senior cybersecurity official from CyberSecurity Malaysia, who requested anonymity. "We cannot afford to outsource our operational sovereignty to foreign AI providers whose guardrails may not reflect our unique risks."

Cisco Talos’ recent evaluation of 66 large language model (LLM) and reasoning combinations found no clear "winner" for security operations. Instead, the study revealed a complex trade-off between efficacy, speed, cost, and consistency. Higher reasoning effort does not necessarily improve analysis — and can degrade performance. Some models took up to 30 minutes to analyze a single log, while others failed to format outputs correctly.

"Choosing an AI model based solely on public leaderboard scores is a recipe for operational disaster," Bianco cautions. He recommends organizations test models against their specific workflows using representative cases, tracking metrics such as response quality, cost, time, consistency, and usable-answer rates.

---

Regional Cyber Threats Intensify: Banking Trojans, Botnets, and Phishing Surge

The call for operational sovereignty in AI security coincides with a rise in sophisticated cyber threats across Southeast Asia. Among the most concerning developments is the evolution of the ToxicPanda banking trojan, now in its 2.0 version. According to Dark Reading, the malware has expanded its arsenal by 167 remote commands and broadened its targeting from 16 to 349 financial institutions, including banks, e-wallets, and cryptocurrency platforms.

ToxicPanda’s expansion highlights the growing sophistication of cybercriminals targeting Malaysia’s rapidly digitizing financial sector. The central bank, Bank Negara Malaysia, has previously warned of rising malware attacks on mobile banking and digital payment systems.

Meanwhile, Interpol’s recent Operation Jackal IV saw law enforcement agencies from 22 countries, including Malaysia, arrest 58 suspects and identify 263 more in a coordinated crackdown on West African cybercrime networks. The operation, spanning six continents, follows two earlier phases in 2022 and 2023 that resulted in approximately 200 arrests and the seizure of millions of dollars in illicit assets.

---

New Threats Emerge: Malware Targets Car Infotainment Systems

In a development with regional implications, researchers have identified what appears to be the first malware designed specifically for car head units. The malicious code, linked to the notorious BadBox botnet, was found on an Android-powered infotainment system manufactured by China-based DoFun, a company whose products are widely used across Asia-Pacific, including Malaysia.

SecurityWeek reports that the malware turns infected head units into nodes in a botnet, potentially enabling large-scale distributed denial-of-service (DDoS) attacks or data exfiltration. The discovery underscores the vulnerabilities in connected vehicles, a sector Malaysia is actively promoting under its National Automotive Policy.

"As vehicles become more connected, they also become more exposed to cyber threats," said a spokesperson for the Malaysian Automotive Association. "We are working with manufacturers to ensure cybersecurity standards are in place."

---

Critical Infrastructure at Risk: CISA Red Team Exposes Weak Defenses

A recent assessment by the U.S. Cybersecurity and Infrastructure Security Agency (CISA) revealed stark disparities in how organizations detect and respond to cyber intrusions. In two red team exercises targeting critical infrastructure, one organization failed to detect or contain the attack entirely, while another rapidly isolated compromised systems and forced the adversary into an "assume breach" posture.

The findings highlight the importance of continuous monitoring, segmentation, and rapid incident response — capabilities that AI tools must support, not hinder. "Defenders must retain full control over their security processes," Bianco emphasizes. "When guardrails are imposed externally, they can become barriers to effective defense."

---

Phishing Evolves: NovaCookies Abuses Docusign to Steal Microsoft 365 Sessions

A new phishing campaign, tracked as NovaCookies, is leveraging legitimate Docusign notifications to steal Microsoft 365 session tokens. The $320-per-month phishing-as-a-service platform has been used to target hundreds of organizations across the U.S., U.K., Canada, Germany, and Australia. While Malaysia has not been named among the targeted regions, the campaign’s use of trusted services like Docusign poses a significant risk to local enterprises.

Security researchers warn that such attacks are difficult to detect, as they abuse authentic email workflows and cloud services. Organizations are advised to implement multi-factor authentication (MFA) and monitor for anomalous session activity.

---

Malaysian Firms Urged to Test AI Models Rigorously Before Deployment

Cisco Talos’ evaluation methodology offers a practical framework for Malaysian organizations seeking to deploy AI in security operations. Key recommendations include:

- Building a focused set of representative test cases. - Using actual prompts and tools analysts will employ. - Tracking performance metrics in a simple spreadsheet. - Establishing acceptable thresholds for quality, cost, time, and consistency. - Regularly revisiting model choices as AI technology and pricing evolve.

"Malaysian companies cannot afford to treat AI as a black box," said a cybersecurity consultant based in Kuala Lumpur. "Local threat actors are becoming more sophisticated. Our defenses must be equally adaptive and in our control."

---

Forward Look: Balancing Innovation and Sovereignty in AI Security

As Malaysia positions itself as a regional leader in digital economy growth, the tension between innovation and security control has never been more pronounced. The rise of agentic AI in cybersecurity offers immense potential but also introduces new risks if guardrails are not carefully designed and controlled.

Industry observers say the solution lies in a hybrid approach: leveraging advanced AI capabilities while maintaining operational sovereignty. This means customizing guardrails to local threat models, ensuring flexibility for authorized interventions, and rigorously testing models before deployment.

"The defender’s advantage is not guaranteed — it must be actively preserved," Bianco concludes. "In the age of AI, that advantage depends on who controls the guardrails."

Related: Cisco Talos · CyberSecurity Malaysia · David Bianco · Kuala Lumpur