Android malware that acts like a person and AI agents that act like malware

The security landscape is rapidly evolving with the advent of sophisticated AI, creating new challenges for defenders. Attackers are leveraging legitimate tools for illegitimate ends, exploiting the gap between technological deployment and governance, and refining malware to bypass modern defenses.

The New Frontier: Malicious AI Agents

We are beginning to see proofs-of-concept for weaponized AI agents, signaling a new class of attacks. Two recent examples highlight the emerging threats:

• “Kofish” (Datadog): This technique uses Microsoft Copilot Studio to build malicious AI agents that covertly steal OAuth tokens, preying on users’ implicit trust in platforms hosted by major tech companies.
• “Agent Session Smuggling” (Palo Alto): This method uses the agent-to-agent communication protocol to secretly transmit malicious commands. It effectively allows a malicious agent to “socially engineer” a legitimate agent without the user’s knowledge.

These developments underscore a critical point: the autonomous and non-deterministic nature of AI agents creates significant oversight challenges. Attackers will increasingly exploit these characteristics.

Rethinking Identity for Non-Human Actors

A recurring theme is that many of these new attacks are extensions of existing credential-based attacks. Just as attackers target valid credentials for human employees, they are now targeting AI agents, which can be considered the “next level of insiders.”

This necessitates a fundamental shift in Identity and Access Management (IAM):

• Treat Agents Like Insiders: AI agents must be uniquely identified, authenticated, and authorized, similar to human users.
• Fine-Grained Scoping: To prevent coercion, agents must be constrained with “blinders”—strict, fine-grained controls over their permissions based on time, resources, actions, and scope.
• Scale and Automation: The sheer scale of non-human identities means traditional IAM tools are insufficient. The new approach must be automated, dynamic, and proactive.

The Persistent AI Governance Gap

There is a significant disparity between the adoption of AI and the implementation of governance for it. IBM research shows that while 72% of businesses have integrated AI, only about 24% have extensive governance frameworks.

This is a movie we’ve seen before, most recently with the cloud playbook: deploy fast, govern later, get breached in between.

Innovation, driven by business needs for productivity and efficiency, consistently outpaces risk management. The solution is not to ban new technology but to foster a culture of secure enablement. This requires a shift from security as a “department of no” to a shared responsibility, where security provides “guardrails” that allow for speed and innovation within safe boundaries, rather than acting as restrictive “checkpoints.”

The Humanization of Malware

A newly discovered banking trojan, nicknamed Herodotus, evades behavioral detection systems by mimicking human typing. Instead of pasting text instantly, it inputs characters one by one with randomized delays.

While the technique itself is simple, it highlights an arms race in threat detection. Current defenses may rely on simple heuristics like typing speed, but attackers are evolving. We can expect malware to incorporate more dimensions of human behavior to appear legitimate. As one expert noted, attackers are targeting the “low-hanging fruit”—organizations that have not yet adopted multi-dimensional risk analysis that evaluates a combination of factors like user, action, device, network, and transaction details.

Social Engineering at Market Scale

A sophisticated smishing campaign is manipulating stock prices. Attackers steal brokerage account credentials and then:

1. Liquidate the victim’s existing investments.
2. Use the funds to buy large amounts of low-liquidity stocks (e.g., penny stocks), artificially inflating the price.
3. Sell their own holdings of that stock at the inflated price for a significant profit.

While the market manipulation aspect is novel, the core of the attack remains a classic credential theft problem. It demonstrates how attackers are finding new ways to monetize compromised accounts with a higher and faster return on investment than traditional methods like ransomware. This underscores the need for multi-layered defenses beyond just MFA, incorporating behavioral analytics to detect anomalous activity.

The State of Bug Bounties

Bug bounty programs are paying out record sums, with platforms like HackerOne hitting all-time highs. This trend is driven by two key factors:

• For Companies: Paying a large bounty is a form of insurance—a million-dollar payout is preferable to a ten-million-dollar ransomware demand.
• For Hackers: It provides a legal and lucrative path for security researchers to apply their skills.

However, bug bounties are not a panacea; they are one tool in a comprehensive security program, often serving as the last line of defense. The largest bounties are reserved for highly complex, hard-to-find vulnerabilities that nation-state actors would otherwise purchase.

In the short term, human ingenuity remains essential for finding novel bugs and complex exploit chains that AI cannot. However, as AI automates the discovery of low-level vulnerabilities, a potential long-term risk is the shrinking pipeline of human experts capable of performing this high-level, creative security research.

• Tags:
• AI Agents
• AI Governance
• Feature
• IAM
• Social Engineering
• Malware
• Smishing
• Bug Bounty