Laporan Akamai: Agen AI Tanpa Pengawasan Jadi Ancaman Siber Baru bagi Bisnis

Jakarta, CMNexus. Kecerdasan buatan (AI) berbasis agen atau agentic AI kini dipakai banyak bisnis untuk menjalankan pekerjaan nyata: membalas pesan pelanggan, menyusun pesanan, memindahkan data antar aplikasi, hingga menjadwalkan layanan. Laporan terbaru Akamai menunjukkan lompatan efisiensi itu membawa risiko baru yang jarang muncul di laporan keuangan pemilik usaha.

Dari asisten yang menjawab menjadi agen yang bertindak

Laporan State of the Internet (SOTI) Security bertajuk Speed, Scale, and Nonhuman Identity: The Agentic Threat Landscape menyebut cara mengamankan bisnis modern telah bergeser: dari mengelola identitas dan hak akses manusia menjadi mengawasi perilaku entitas non-manusia.

"AI menghadirkan situasi di mana banyak ancaman muncul bersamaan bagi ekosistem keamanan," kata Chief Security Officer Akamai, Boaz Gelbord, dikutip Liputan6 (27 September 2026).

Akamai menyoroti dua kesulitan utama: memasang pengaman (guardrail) bagi agen otonom, dan kecepatan model AI menemukan celah keamanan yang sering melampaui siklus penambalan (patching) berbasis manusia.

Empat angka yang perlu dibaca pemilik usaha

47 persen interaksi chatbot AI di lingkungan perusahaan berjalan melalui akun pribadi yang tidak terkelola, sehingga aktivitas itu luput dari pengawasan tim TI (Akamai, September 2026).
Lebih dari 6 persen percakapan chatbot AI perusahaan memuat data sensitif, terutama informasi identitas pribadi (Akamai).
Lebih dari 40 persen pengguna korporat memasang ekstensi browser berbasis AI, dan seperempat dari ekstensi itu mengubah izin akses dalam 12 bulan (Akamai).
Pemimpin keamanan memperkirakan agen AI yang beroperasi tanpa pengawasan ketat akan menjadi salah satu ancaman siber paling berbahaya sebelum 2030 (Akamai, dikutip Liputan6).

Malware berkedok "aplikasi AI gratis" menyasar usaha kecil

Risiko tidak hanya datang dari dalam. Kaspersky mencatat lebih dari 33.300 serangan terhadap usaha kecil dan menengah pada Januari–April 2026 berupa program berbahaya yang menyamar sebagai layanan AI populer. Angka itu hampir lima kali lipat dibanding periode yang sama pada 2025.

Umpan yang paling sering dipakai adalah aplikasi palsu yang mengaku sebagai ChatGPT (42 persen), Claude (24 persen), dan DeepSeek (20 persen). Kaspersky juga menemukan lebih dari 1.100 sampel berbahaya unik dari lima layanan AI yang paling sering ditiru, naik 21 persen dibanding tahun sebelumnya.

"Lanskap ancaman terus berkembang dengan umpan baru yang muncul," kata pakar keamanan Kaspersky, Vasily Kolesnikov. Ia menyebut pada empat bulan pertama 2026 timnya mendeteksi ratusan serangan yang menyamar sebagai OpenClaw, alat AI yang popularitasnya melesat sepanjang 2026.

Sebagai pembanding, serangan berkedok aplikasi pesan dan konferensi video seperti Telegram, WhatsApp, Zoom, dan Microsoft Teams bahkan lebih banyak: hampir 415.000 pada periode yang sama.

Titik lemah yang jarang diawasi: protokol penghubung agen

Agen AI bisa bekerja karena terhubung ke banyak sistem lewat Model Context Protocol (MCP), standar terbuka yang membuat agen dapat memanggil alat dan menjalankan tindakan lintas aplikasi. Akamai menemukan pengawasan MCP justru berada di urutan paling bawah daftar prioritas keamanan perusahaan, sehingga muncul kesenjangan visibilitas yang tajam.

Padahal lembaga riset Cloud Security Alliance (CSA) Labs mencatat 1.862 server MCP yang terbuka di internet dan melayani permintaan tanpa autentikasi pada pemindaian 2025 (catatan riset Mei 2026). Analisis Docker terhadap ekosistem MCP menemukan kelemahan command injection pada 43 persen server yang diperiksa. Panduan keamanan MCP dari NSA/CISA (Juni 2026) juga menegaskan banyak implementasi mengabaikan autentikasi dan kontrol akses.

Enam langkah praktis untuk pemilik usaha

Mulai dari akses seperlunya. Beri agen AI izin seminimal mungkin; jangan sekaligus ke seluruh percakapan pelanggan, data keuangan, dan berkas di komputer.
Pasang titik persetujuan manusia untuk tindakan yang sulit dibatalkan: transfer dana, hapus data, kirim pesan massal, atau ubah harga.
Unduh hanya dari situs resmi. Sebelum memasang aplikasi AI, periksa ejaan alamat situs dan tautan di email, karena malware menyamar dengan nama yang mirip.
Pakai akun perusahaan, bukan akun pribadi karyawan. Percakapan kerja di akun pribadi tidak terekam dan memudahkan data usaha keluar tanpa jejak.
Audit ekstensi browser dan server MCP secara berkala, dan simpan catatan setiap perintah yang dijalankan agen.
Latih karyawan soal prompt injection dan aplikasi palsu; satu unduhan dari email penipuan bisa membuka jalan masuk serangan lanjutan berbulan-bulan kemudian.

Yang perlu diingat

Agen AI bukan sekadar alat produktivitas baru. Ia bekerja dengan kecepatan otomatisasi namun membawa wewenang yang mirip seorang karyawan. Usaha kecil yang menetapkan batas akses, titik persetujuan manusia, dan catatan aktivitas sejak awal bisa memetik manfaatnya tanpa membayar mahal di kemudian hari.

Baca juga berita teknologi dan UMKM lain di cmnexus.com/news.

Sumber

Akamai — Akamai Report: Securing Agentic AI Requires Shift to Behavioral Governance (22 September 2026) — https://www.akamai.com/newsroom/press-release/akamai-report-securing-agentic-ai-requires-shift-to-behavioral-governance
Akamai — Speed, Scale, and Nonhuman Identity: The Agentic Threat Landscape (SOTI Security Report) — https://www.akamai.com/lp/state-of-the-internet/agentic-threat-landscape-report
Liputan6 — Ketika Agen AI Mulai Bertindak Sendiri Tanpa Pengawasan (27 September 2026) — https://www.liputan6.com/tekno/read/8300185/ketika-agen-ai-mulai-bertindak-sendiri-tanpa-pengawasan
Kaspersky — Malware attacks on SMBs disguised as AI services surged by five times in 2026 (25 Juni 2026) — https://www.kaspersky.co.uk/about/press-releases/malware-attacks-on-smbs-disguised-as-ai-services-surged-by-five-times-in-2026-kaspersky-reports
Manila Bulletin — Malware attacks on SMBs disguised as AI services surged by five times in 2026, Kaspersky reports (30 Juli 2026) — https://mb.com.ph/2026/07/30/malware-attacks-on-smbs-disguised-as-ai-services-surged-by-five-times-in-2026-kaspersky-reports
NSA/CISA — Model Context Protocol (MCP): Security Design Considerations (Juni 2026) — https://media.defense.gov/2026/Jun/02/2003943289/-1/-1/0/CSI_MCP_SECURITY.PDF
Docker — MCP Security: Risks, Challenges, and How to Mitigate — https://www.docker.com/blog/mcp-security-explained/
Cloud Security Alliance Labs — MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure (4 Mei 2026) — https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-security-crisis-20260504-csa-styled/

Jakarta, CMNexus. Agent-based artificial intelligence, or agentic AI, is now deployed by many businesses to perform real-world tasks: responding to customer inquiries, drafting orders, transferring data between applications, and scheduling services. Akamai's latest report indicates that this leap in efficiency brings new risks that rarely appear on a business owner's balance sheet.

From Assistants That Answer to Agents That Act

The State of the Internet (SOTI) Security report titled Speed, Scale, and Nonhuman Identity: The Agentic Threat Landscape notes that securing modern businesses has shifted: from managing human identities and access permissions to monitoring non-human entity behavior.

"AI presents a situation where multiple threats converge simultaneously for the security ecosystem," said Akamai Chief Security Officer Boaz Gelbord, quoted by Liputan6 (27 September 2026).

Akamai highlighted two primary challenges: establishing guardrails for autonomous agents, and the speed at which AI models uncover security flaws, which often outpaces human-driven patching cycles.

Four Numbers Every Business Owner Needs to Know

47 percent of AI chatbot interactions in corporate environments take place through unmanaged personal accounts, leaving these activities outside the oversight of IT teams (Akamai, September 2026).
Over 6 percent of enterprise AI chatbot conversations contain sensitive data, particularly personally identifiable information (Akamai).
Over 40 percent of enterprise users have installed AI-powered browser extensions, and a quarter of those extensions altered access permissions within 12 months (Akamai).
Security leaders anticipate that AI agents operating without strict supervision will become one of the most dangerous cyber threats before 2030 (Akamai, cited by Liputan6).

Malware Disguised as "Free AI Apps" Targets Small Businesses

The risk does not only originate from within. Kaspersky recorded over 33,300 attacks targeting small and medium-sized businesses between January and April 2026 involving malware masquerading as popular AI services. That figure is nearly five times higher compared to the same period in 2025.

The most frequently used lures were fake applications impersonating ChatGPT (42 percent), Claude (24 percent), and DeepSeek (20 percent). Kaspersky also identified more than 1,100 unique malicious samples across the five most frequently spoofed AI services, representing a 21 percent increase year-over-year.

"The threat landscape continues to evolve with new lures emerging," said Kaspersky security expert Vasily Kolesnikov. He noted that during the first four months of 2026, his team detected hundreds of attacks masquerading as OpenClaw, an AI tool whose popularity surged throughout 2026.

By comparison, attacks disguised as messaging and video conferencing tools such as Telegram, WhatsApp, Zoom, and Microsoft Teams were even more numerous: nearly 415,000 during the same period.

A Rarely Monitored Blind Spot: Agent Communication Protocols

AI agents are able to function because they connect across multiple systems via the Model Context Protocol (MCP), an open standard enabling agents to call tools and execute actions across applications. Akamai found that MCP governance sits at the very bottom of enterprise security priority lists, creating a severe visibility gap.

This comes despite research organization Cloud Security Alliance (CSA) Labs identifying 1,862 MCP servers exposed to the public internet that serviced requests without authentication during a 2025 scan (research note May 2026). A Docker analysis of the MCP ecosystem discovered command injection vulnerabilities in 43 percent of inspected servers. MCP security guidance from NSA/CISA (June 2026) also emphasized that many implementations omit authentication and access controls.

Six Practical Steps for Business Owners

Start with least privilege. Grant AI agents only the minimum necessary permissions; do not immediately grant access to all customer conversations, financial records, and computer files.
Implement human approval checkpoints for actions that are difficult to undo: fund transfers, data deletion, bulk messaging, or price modifications.
Download exclusively from official websites. Before installing an AI application, check the spelling of website addresses and email links, as malware frequently mimics legitimate brand names with slight variations.
Use corporate accounts rather than employee personal accounts. Work-related discussions on personal accounts are not logged, making it easy for corporate data to leak without a trace.
Regularly audit browser extensions and MCP servers, and maintain detailed logs of every command executed by agents.
Train employees on prompt injection and counterfeit applications; a single download from a deceptive email can provide initial access for follow-on attacks months later.

Key Takeaway

AI agents are not merely new productivity tools. They operate with the speed of automation while carrying authority akin to an employee. Small businesses that establish access limits, human checkpoints, and activity logs from the outset can reap the benefits without paying a heavy price down the line.

Read more technology and MSME news at cmnexus.com/news.

Sources

Akamai — Akamai Report: Securing Agentic AI Requires Shift to Behavioral Governance (22 September 2026) — https://www.akamai.com/newsroom/press-release/akamai-report-securing-agentic-ai-requires-shift-to-behavioral-governance
Akamai — Speed, Scale, and Nonhuman Identity: The Agentic Threat Landscape (SOTI Security Report) — https://www.akamai.com/lp/state-of-the-internet/agentic-threat-landscape-report
Liputan6 — Ketika Agen AI Mulai Bertindak Sendiri Tanpa Pengawasan (27 September 2026) — https://www.liputan6.com/tekno/read/8300185/ketika-agen-ai-mulai-bertindak-sendiri-tanpa-pengawasan
Kaspersky — Malware attacks on SMBs disguised as AI services surged by five times in 2026 (25 Juni 2026) — https://www.kaspersky.co.uk/about/press-releases/malware-attacks-on-smbs-disguised-as-ai-services-surged-by-five-times-in-2026-kaspersky-reports
Manila Bulletin — Malware attacks on SMBs disguised as AI services surged by five times in 2026, Kaspersky reports (30 Juli 2026) — https://mb.com.ph/2026/07/30/malware-attacks-on-smbs-disguised-as-ai-services-surged-by-five-times-in-2026-kaspersky-reports
NSA/CISA — Model Context Protocol (MCP): Security Design Considerations (Juni 2026) — https://media.defense.gov/2026/Jun/02/2003943289/-1/-1/0/CSI_MCP_SECURITY.PDF
Docker — MCP Security: Risks, Challenges, and How to Mitigate — https://www.docker.com/blog/mcp-security-explained/
Cloud Security Alliance Labs — MCP Security Crisis: Systemic Design Flaws in AI Agent Infrastructure (4 Mei 2026) — https://labs.cloudsecurityalliance.org/research/csa-research-note-mcp-security-crisis-20260504-csa-styled/