Can an AI agent already replace an L1 analyst?

There is traction in triage and enrichment, but Bengio reports a gap between pre-deployment and real-world performance. Agents still make basic errors. Human supervision remains the operational baseline.

What new attack surface do AI agents bring?

Prompt injections in sources an agent processes automatically: SIEM logs, DNS requests, infrastructure data, shared files. If the agent acts without gating, the attacker shifts from exploiting code to exploiting context.

Is Bengio's report relevant to a corporate board?

Yes. It is the most recent intergovernmental synthesis, with EU, OECD and UN endorsement. It does not set obligations, but it structures the conversation around malicious use, malfunction and systemic risk in terms the board can use to prioritize AI investment and governance.

AI agents in cyberattacks 2026: statistics for SOCs

2026 figures on AI agents used for attack and defense, with the SOC-CMM reading for detection, response and safeguard governance.

Headline figure

77%

of real software vulnerabilities identified by an AI agent in a controlled competition

Source: International AI Safety Report 2026 (DSIT 2026/001, chaired by Yoshua Bengio, more than 100 experts from more than 30 countries, February 2026); cross-checked with Fujitsu 2026 Cybersecurity Predictions and S2 Grupo Cybersecurity Trends 2026.

Executive reading

The 2026 International AI Safety Report, chaired by Yoshua Bengio with backing from the EU, OECD and UN, documents two practical findings for a SOC: AI agents already identify a very high fraction of real software vulnerabilities, and criminal groups plus state-linked actors are using general-purpose AI in their operations. The defense-attack boundary stops being hypothetical. Fujitsu adds a concrete vector: in 2026, prompt injections will appear in unexpected sources like SIEM logs, DNS or infrastructure data that agents process automatically.

For the SOC the implication is twofold. Defensively, AI in triage and enrichment has traction, but the report warns of a gap between pre-deployment and real-world performance. Offensively, SIEM, DNS and infrastructure feeds must be treated as injection surfaces. SOC-CMM measures both pieces: misuse detection capability in Services, safeguard governance in Process, telemetry monitoring in Technology.

Report data

Metric	Value	Reading
Real vulnerabilities found by an AI agent in competition	77%	Bengio 2026.
Daily general-purpose AI users (global estimate)	~1 B	Bengio 2026.
Companies that published Frontier AI Safety Frameworks in 2025	12	Bengio 2026.
Data breaches related to ransomware in 2025 (estimate)	44%	S2 Grupo 2026.
Cyberattacks directly or indirectly involving employees	88%	S2 Grupo 2026.
Share of global ransomware victims in European organizations (2025)	22%	S2 Grupo 2026.
Time from initial breach to ransomware execution (observed)	<24 h	S2 Grupo 2026.
Trust in AI handling core functions without supervision	42%	Fortinet 2026.

What to do with this data

Treat SIEM, DNS and vendor feeds as potential prompt injection surfaces.
Define guardrails and human approval for high-risk automated actions.
Measure the gap between pre-deployment evaluation and production behavior.
Inventory production agents and operational dependency by process.
Document the telemetry chain and AI decision traceability for audit.

FAQ

All SOC-CMM statistics Independent SOC-CMM assessment SOC-CMM AI assurance