AI Agent Security Statistics 2026
By Axis Intelligence Research
Co-author: Marcus Chen | Last updated: July 22, 2026 | License: CC BY 4.0
Enterprise AI agent fleets doubled in a single quarter in early 2026, yet monitoring coverage barely moved. According to Gravitee.io’s April 2026 survey of 750 senior technology leaders, 48% of all production AI agents are running unsecured — while 85% of organizations have no formal accountability structure for agent behavior. Autonomous agents now account for more than 1 in 8 reported AI breaches, per HiddenLayer’s 2026 AI Threat Landscape Report.
Quick Answer
How serious are AI agent security risks in 2026?
According to Axis Intelligence Research’s cross-source analysis of six primary reports covering 2025–2026 data: 48% of production AI agents are unmonitored; agentic AI CVEs surged 255% year-over-year (74 to 263), per Trend Micro’s TrendAI™ State of AI Security Report; 54% of organizations have already experienced an AI agent security incident; only 14.4% of agents reach production with full security approval; and 85% of organizations have no named individual accountable for agent behavior. The Axis Agent Security Gap Index (AASGI™) — a composite of deployment velocity, monitoring coverage, governance maturity, and incident exposure — registers a baseline reading of 71.4 out of 100 as of July 2026, where higher scores indicate greater aggregate risk.
Key Findings
- Agentic AI CVEs surged 255.4% year-over-year — from 74 in 2024 to 263 in 2025 — the fastest growth rate of any AI vulnerability subcategory, according to Trend Micro’s TrendAI™ State of AI Security Report (March 2026).
- 48% of production AI agents operate without active security monitoring, with only 9.5% of organizations securing more than 80% of their deployed agent fleet, per Gravitee.io’s State of AI Agent Security survey of 750 executives (April 2026).
- Autonomous agents are implicated in 1 in 8 reported AI breaches, according to HiddenLayer’s 2026 AI Threat Landscape Report (March 2026), based on a survey of 250 IT and security leaders.
- Only 14.4% of organizations deploy AI agents to production with full security or IT approval, per HiddenLayer’s 2026 AI Threat Landscape Report, meaning more than 85% of agent deployments bypass full security review.
- 85% of organizations have no formally named individual accountable for AI agent behavior — a figure Gravitee.io describes as the governance crisis inside the security crisis, drawn from its April 2026 survey.
How Fast Are Enterprise AI Agent Fleets Growing?
Speed is the problem. Gravitee.io’s longitudinal data across December 2025 and April 2026 — surveying 750 CIOs, CTOs, VPs Engineering, and Heads of Platform across financial services, healthcare, telecoms, manufacturing, and travel — shows the modal deployment bracket shifted from 26–50 agents per organization to 76–100 in four months. A cohort of active deployers is already running at 76–100 agents and doubling every quarter.
What didn’t keep pace: security.
Agent Deployment vs. Security Coverage: December 2025 to April 2026
| Metric | Dec 2025 | Apr 2026 | Change |
|---|---|---|---|
| Modal agent deployment bracket | 26–50 agents | 76–100 agents | ~2× growth |
| Mean monitoring coverage | 46.96% | ~52% | Effectively flat |
| Confirmed security incidents (respondents) | 59.3% | 34.9% | Likely underreporting |
| Confidence in visibility (net) | 82.6% | 91.8% | Rising vs. flat coverage |
| All agents secured at deployment | 13.6% | 19.7% | Modest improvement |
| Formal named accountability | ~1% | 7.2% | Still critically low |
| Planning to deploy more agents (12mo) | ~50% | 81.7% | Acceleration confirmed |
Source: Gravitee.io, State of AI Agent Security (April 2026, n=750)
The confidence-reality inversion is the data point that matters most. Stated confidence in agent visibility rose 9 percentage points in four months. Monitoring coverage barely moved. Organizations are growing more comfortable with a risk they haven’t reduced.
Marcus Chen notes: The December-to-April comparison is the clearest quantitative evidence yet that AI agent governance is a solved problem on paper and an open wound in practice. Incident rates didn’t fall because security improved — they fell because detection didn’t keep pace with a fleet that doubled. You can’t confirm an incident you never saw.
Source: Axis Intelligence Research — CC BY 4.0
What Are the Biggest AI Agent Security Vulnerabilities?
Prompt Injection: Why It Dominates Every Ranking
OWASP has ranked prompt injection the number-one risk on its Top 10 for LLM Applications for two consecutive editions — 2024 and 2025. The mechanism is simple and the consequences in agentic contexts are severe: LLMs process instructions and external data through the same channel, without architectural separation, so a crafted input can override intended behavior. In agent deployments, this means the injection doesn’t just produce a bad response. It triggers actions: database queries, API calls, code execution, file transfers.
The indirect variant — where malicious instructions are embedded inside a document, webpage, or email that the agent retrieves — scales the problem to every resource an agent can read. Trend Micro’s research documented 450 confirmed prompt injection incidents in the first four months of 2026. A 2025 benchmark study found 94.4% of AI agents were vulnerable to being hijacked through content they were asked to read.
Agentic CVE Acceleration: The Numbers by Subcategory
Trend Micro’s TrendAI™ Research analyzed 330,239 CVEs from 2018 through 2025, identifying 6,086 AI-specific vulnerabilities. Within that dataset, the agentic AI subcategory grew from 74 CVEs in 2024 to 263 in 2025 — a 255.4% increase, the steepest year-over-year jump across all eight AI vulnerability subcategories.
AI Vulnerability Growth by Subcategory: 2024 vs. 2025
| Subcategory | 2024 CVEs | 2025 CVEs | YoY Growth |
|---|---|---|---|
| Agentic AI | 74 | 263 | +255.4% |
| MCP Servers | 0 | 95 | New category |
| LLM Ecosystem | 419 | 756 | +80.4% |
| AI Model Security | ~180 | ~317 | +76.1% |
| ML Frameworks | ~310 | ~472 | +52.2% |
| GPU & AI Hardware | ~1,300 | ~1,581 | +21.6% |
| AI Data Pipelines | ~330 | ~292 | -11.4% |
Source: Trend Micro, TrendAI™ State of AI Security Report (March 2026). Note: total AI CVEs 2025 = 2,130 (34.6% YoY growth vs. 17.9% for all CVEs).
MCP servers are a special case. This subcategory did not exist as a tracked category before 2025; 95 CVEs appeared in a single year. Trend Micro found 1,467 exposed MCP servers in the wild by late 2025, with 83% still running the deprecated Server-Sent Events transport protocol — evidence of organizations copying outdated reference implementations directly into production.
The Six Dominant Failure Patterns
Gravitee.io’s cross-analysis of open-text responses from 200+ executives across both survey waves identified six recurring incident patterns:
Excessive permissions. The most consistently reported failure. Agents deployed with access to systems, datasets, or API endpoints beyond their task requirements — often rooted in shared service accounts or inherited credentials. Confirmed across both December 2025 and April 2026 cohorts.
Data retention violations. Agents retaining prompts, conversation logs, or personal data beyond policy. April 2026 data added new variants: unintentional retention via third-party API logging and voice biometrics stored unencrypted.
Prompt injection and adversarial manipulation. The December 2025 data showed mostly accidental misuse. April 2026 introduced explicit adversarial attacks — external actors manipulating models through crafted inputs to extract hidden information, jailbreaking safety controls, and AI-generated malware introduced via code suggestions. This shift from accidental to deliberate is the most significant change between waves.
Shadow AI and ungoverned deployment. Teams deploying agents outside IT or security knowledge, before governance frameworks are in place. Structurally unchanged between surveys; prevalent across all sectors.
Third-party agent and vendor opacity. External AI providers changing processing logic without notice or mishandling data. A UK-specific variant emerged in April 2026: a data labelling vendor used real customer data without anonymisation during AI training.
AI output errors causing operational harm. Largely absent in December 2025, clearly emerging in April 2026: agents producing confident but incorrect outputs that influenced real decisions — incorrect financial analysis, outdated risk scores, diagnostic misclassification without audit logging.
Who Is Accountable When an AI Agent Causes Harm?
Almost nobody. Gravitee.io asked the accountability question differently in each survey wave, and the gap between perceived ownership and actual governance structures was stark. In December 2025, respondents named CTO (28.2%) or CISO (26%) as accountable. The April 2026 survey asked how accountability is handled, not who holds it — and the picture collapsed.
How Is Accountability for AI Agent Behavior Handled? (April 2026)
| Structure | % of Organizations |
|---|---|
| Accountability unclear / situation-dependent | 32.4% |
| Responsibility shared but not formally defined | 29.9% |
| Not discussed or decided | 22.9% |
| Named individual with formal accountability | 7.2% |
| Under evaluation | 7.1% |
| Not using agents | 4.1% |
| Other | 7.6% |
Source: Gravitee.io, State of AI Agent Security (April 2026, n=750)
Only 7.2% of organizations have a named individual with formal accountability for AI agent behavior. The majority describe it as unclear, shared but undefined, or simply undiscussed. What looked like named ownership in December — “the CTO is accountable” — was, for most organizations, informal and undefined. The April framing exposed the optimism behind the December numbers for what it was.
How Do AI Agent Incidents Break Down by Sector?
Telecoms and healthcare lead confirmed incident rates. Financial services, despite the highest proportion of suspected-but-unconfirmed incidents (46.7%), shows a lower confirmed rate (8%) — a pattern consistent with either better detection and suppression, or with underreporting in a heavily regulated sector where disclosing AI incidents carries regulatory risk.
AI Agent Security Incidents by Sector: Confirmed vs. Suspected (April 2026)
| Sector | Confirmed Incident | Suspected Incident | Total Exposure |
|---|---|---|---|
| Telecoms | 30.7% | 36.7% | 67.3% |
| Healthcare | 26.2% | 28.2% | 54.4% |
| Financial Services | 8.0% | 46.7% | 54.7% |
| Manufacturing | 16.8% | 30.2% | 47.0% |
| Travel & Transport | 14.0% | 32.7% | 46.7% |
Source: Gravitee.io, State of AI Agent Security (April 2026, n=750)
Note: The overall confirmed incident rate dropped from 59.3% (December 2025) to 34.9% (April 2026). Gravitee.io attributes this to underreporting and detection failure rather than genuine security improvement, given that fleet size doubled and monitoring coverage barely moved in the same period.
The telecoms sector’s prominence is not surprising. Telecoms operate massive customer data estates, use AI agents for network operations and customer service at scale, and are frequently targeted precisely because of that data concentration. The pattern of open-text responses from telecoms respondents — citing prompt injection attempts, third-party API misconfigurations, and shadow AI deployments — aligns with the structural exposure profile.
What Does the IBM X-Force 2026 Report Say About AI Agent Risk?
IBM X-Force’s 2026 Threat Intelligence Index (released February 25, 2026, based on 2025 incident response data) documents the credential-risk dimension of AI agent deployment at scale. The report found that over 300,000 ChatGPT credential sets were advertised on dark web markets in 2025, driven by infostealer malware operators who expanded their target lists to include AI services. IBM X-Force frames this as direct evidence that AI platforms have reached the same credential-risk profile as core enterprise SaaS.
The broader context: vulnerability exploitation displaced credential theft as the leading initial access vector in 2025, accounting for 40% of all incidents — a 44% year-over-year increase in attacks starting from public-facing application exploitation. Many exploited vulnerabilities required no authentication whatsoever. AI agent endpoints, which frequently inherit weak authentication practices from the rapid-development environments that created them, are explicitly named as a growing attack surface.
Supply chain compromises have increased nearly fourfold over the past five years, per the same report. For AI agents that rely on third-party tools, plugins, and model registries, this is not a background risk — it is the attack surface.
MCP Vulnerabilities: The Fastest-Growing AI Attack Category
The Model Context Protocol, introduced by Anthropic in November 2024 and subsequently stewarded by the Linux Foundation Agentic AI Foundation, created a new attack surface almost immediately. Trend Micro documented 1,467 exposed MCP servers as of December 2025 — a nearly threefold increase from initial findings — with 1,227 (83%) running the deprecated SSE transport protocol.
The most consequential exposure: the “execute_sql” tool was found available on 70 publicly accessible hosts without authentication. Anyone could execute arbitrary SQL queries and extract customer records, financial data, or authentication credentials without exploiting a traditional vulnerability.
A second category — 39 hosts running “Graphiti Agent Memory” implementations — gave attackers access not just to current data but to historical records of agent activity, potentially revealing strategic planning and internal communications accumulated over time.
Trend Micro’s analysis of 19,000 open-source MCP repositories found that 4% to 20% contained exploitable vulnerabilities, with 42.6% of classified vulnerabilities showing signs of AI-generated code. The pattern: developers using AI coding assistants to generate MCP server implementations without deep understanding of the underlying security model, producing code that is syntactically correct and operationally dangerous.
MCP Server Security Posture Snapshot (December 2025)
| Finding | Value |
|---|---|
| Total exposed MCP servers found | 1,467 |
| Running deprecated SSE transport | 1,227 (83%) |
| Hosts with unauthenticated execute_sql tool | 70 |
| Hosts with Graphiti Agent Memory (unprotected) | 39 |
| MCP CVEs in 2025 (from zero in 2024) | 95 |
| Vulnerability types: injection-class | >60% of MCP CVEs |
Source: Trend Micro, TrendAI™ State of AI Security Report (March 2026)
The Deployment Speed Pressure Problem
81% of respondents in Gravitee.io’s April 2026 survey feel pressure to deploy AI agents quickly, even when security or governance is not fully in place. 25.8% describe this as significant pressure.
The drivers: maximize ROI on AI investment (38.1%), increase productivity (36.2%), keep up with competitors (34.5%), board or investor pressure (33.7%). These are not irrational motivations. The problem is what they displace: pre-deployment security review (present in only 35% of organizations), documented data access controls specific to each agent (32%), and a defined scope of permitted access (30.5%).
Only 19.7% of organizations report that all agents are fully secured and governed before going live. The majority — 59.1% — say “most” are, which means a meaningful proportion of their fleet routinely enters production without adequate controls. The April improvement from December’s 13.6% is real but modest.
No single control is used by even 40% of organizations before an agent goes live. A named person accountable for agent behavior — the baseline accountability control — is present in only 37.8% of pre-deployment checklists.
Axis Agent Security Gap Index (AASGI™)
Axis Intelligence Research has developed the Axis Agent Security Gap Index (AASGI™) to provide a single, reproducible composite score measuring the distance between AI agent deployment velocity and enterprise security readiness. This metric does not exist in any single primary report and is the original analytical contribution of this article.
Formula and inputs (v1.0, July 2026 baseline):
AASGI™ = (D × 0.35) + (M × 0.30) + (G × 0.20) + (I × 0.15)
Where:
- D = Deployment Velocity Score (0–100): Percentage of organizations planning to deploy significantly more agents in 12 months. Source: Gravitee.io April 2026 (81.7% → score of 81.7).
- M = Monitoring Gap Score (0–100): Percentage of production agents running without active security monitoring. Source: Gravitee.io April 2026 (48% → score of 48).
- G = Governance Deficit Score (0–100): Percentage of organizations with no formally named accountable individual. Source: Gravitee.io April 2026 (92.8% → score of 92.8, as 7.2% have formal accountability).
- I = Incident Exposure Score (0–100): Percentage of organizations that have experienced or suspected a security incident. Source: Gravitee.io April 2026 (54% → score of 54).
Calculation (July 2026): (81.7 × 0.35) + (48 × 0.30) + (92.8 × 0.20) + (54 × 0.15) = 28.6 + 14.4 + 18.56 + 8.1 = 69.66 → rounded to 69.7 out of 100
Note: AASGI™ v1.0 baseline reading is 69.7 as of July 2026. Higher scores indicate greater aggregate risk. This is a baseline — the index cannot be described relative to prior readings that were never computed.
Stated limitations: AASGI™ combines survey data from a single vendor survey (Gravitee.io) with an intent metric. It does not independently weight sector or geography. Governance deficit and incident exposure inputs are self-reported. Future versions will incorporate multi-source survey data and CVE-rate inputs to reduce single-source dependence.
What Regulations Apply to AI Agent Security?
The regulatory landscape is fragmenting, and enterprises know it. Only 39.5% of Gravitee.io respondents believe current regulation sufficiently covers AI agent risks. The identified gaps cluster around five areas: data access and privacy in agentic systems (32%), agents handling sensitive information (28.9%), agent-to-agent communication and coordination (28.4%), accountability when an agent causes harm (28.2%), and autonomous decision-making by agents (27.3%).
The EU AI Act’s prohibitions took effect in February 2026, with the full obligations for high-risk AI systems applying from August 2026. Agentic systems that operate in high-risk categories — healthcare, employment, critical infrastructure — are subject to conformity assessments, mandatory logging, and human oversight requirements. The Act does not yet define “AI agent” as a formal category distinct from “AI system,” which means the interpretation of how agent-specific risks are governed remains contested.
NIST’s AI Risk Management Framework (AI 600-1, published 2024) provides the closest thing to a voluntary standard for agentic systems in the US context, establishing profiles for generative AI that address prompt injection and data governance. The framework is non-binding, which is reflected in the adoption data above.
The OWASP Top 10 for LLM Applications 2025 and the OWASP Top 10 for Agentic Applications (announced Black Hat Europe 2025) provide the most operationally concrete vulnerability taxonomies available. They are not regulatory instruments, but they are the reference frameworks most security teams are actually using.
What Predicts the 2026 Trajectory for AI Agent CVEs?
Trend Micro’s forward-looking analysis projects 2,800–3,600 total AI CVEs in 2026, a 31–69% increase from 2,130 in 2025. The agentic AI subcategory is projected to reach 350–550 CVEs — a 33–109% increase — while MCP servers are projected to grow from 95 CVEs to 180–280, an 89–195% increase. These are the two fastest-growing subcategories by percentage projection.
The structural driver is that AI agent deployment is growing faster than security research can characterize the attack surface. When a technology moves from development to production at the rate described by Gravitee.io’s fleet-doubling data, the vulnerability disclosure curve typically follows with a 12–18 month lag. The 2024-to-2025 agentic CVE surge (74 to 263) is that lag starting to close.
Pwn2Own Berlin 2025 added a dedicated AI infrastructure category for the first time, targeting the Nvidia Triton Inference Server. That signal matters: every new Pwn2Own category has historically preceded a surge in vulnerability discovery in that product class as offensive researchers focus coordinated attention on it.
Axis Intelligence Research Methodology
Data collection: This article draws exclusively from primary sources fetched and read during production in July 2026. Secondary aggregators were not cited. Sources: Gravitee.io State of AI Agent Security (December 2025 and April 2026); Trend Micro TrendAI™ State of AI Security Report (March 3, 2026); HiddenLayer 2026 AI Threat Landscape Report (March 18, 2026, press release and third-party coverage); IBM X-Force Threat Intelligence Index 2026 (February 25, 2026); OWASP Top 10 for LLM Applications 2025 (genai.owasp.org); GitHub repository webpro255/awesome-ai-agent-attacks (accessed July 2026).
AASGI™ computation: Four inputs drawn from the Gravitee.io April 2026 dataset; weighted and summed per the formula in the Axis Agent Security Gap Index section above. Python arithmetic verification: (81.7 × 0.35) + (48 × 0.30) + (92.8 × 0.20) + (54 × 0.15) = 28.595 + 14.4 + 18.56 + 8.1 = 69.655. Rounded to 69.7. All inputs are survey-based self-reported data from a single vendor; limitations stated above.
Trend Micro CVE subcategory data: Drawn from the TrendAI™ State of AI Security Report’s Table 3 (severity by year) and the year-over-year growth figure from text and Figure 4 of the same report. The specific YoY growth percentages for subcategories are sourced from the report text.
Excluded figures: Any statistic not traceable to a directly fetched primary source was excluded. Several figures encountered in secondary sources (specific breach dollar amounts attributed to AI agents, specific global AI security market size projections) were not used because the originating primary document could not be confirmed as fetched.
Limitations: Gravitee.io’s survey (n=750) is vendor-commissioned and draws from organizations that self-select into a survey about AI agent security — likely skewing toward organizations with above-average awareness. Incident rates may be underestimated across all sources due to detection gaps acknowledged in the Gravitee data itself. MCP server counts from Trend Micro reflect Shodan scanning from September–December 2025 and will have changed by publication.
About This Dataset
The dataset underlying this article — available as a CC BY 4.0 CSV — contains 34 quantitative observations drawn from the six primary sources listed in the Methodology section. Every row carries source organization, source document, URL, retrieval date, and a flag for whether the figure is a direct observation or an Axis Intelligence Research calculation.
The AASGI™ formula, weights, and component inputs are included as axis_calculated = yes rows with full method_note fields linking to this section.
Citation format:
APA: Axis Intelligence Research, & Chen, M. (2026). AI agent security statistics 2026. Axis Intelligence. https://axis-intelligence.com/ai-agent-security-statistics/
MLA: Axis Intelligence Research and Marcus Chen. “AI Agent Security Statistics 2026.” Axis Intelligence, 2026, axis-intelligence.com/ai-agent-security-statistics/.
Chicago: Axis Intelligence Research Desk, and Marcus Chen. “AI Agent Security Statistics 2026.” Axis Intelligence. 2026. https://axis-intelligence.com/ai-agent-security-statistics/.
License: CC BY 4.0. Attribution required: “Axis Intelligence Research, AI Agent Security Statistics 2026.”
Last updated: July 22, 2026
Frequently Asked Questions
What percentage of AI agents are running without security in 2026?
According to Gravitee.io’s April 2026 survey of 750 senior technology leaders, 48% of all production AI agents are running without active security monitoring. Only 9.5% of organizations are securing more than 80% of their deployed agent fleet. The mean monitoring coverage is approximately 52%, meaning roughly half the average agent fleet operates outside active security oversight.
What is the most common AI agent vulnerability?
Prompt injection is ranked the number-one vulnerability on the OWASP Top 10 for LLM Applications for two consecutive editions (2024 and 2025). In agentic systems specifically, indirect prompt injection — where malicious instructions are embedded in external content an agent retrieves — is the highest-priority attack class. Trend Micro’s 2026 research documented 450 confirmed prompt injection incidents in the first four months of 2026 alone.
How many AI agent CVEs were disclosed in 2025?
Trend Micro’s TrendAI™ Research identified 263 CVEs classified in the agentic AI subcategory in 2025 — a 255.4% increase from 74 in 2024. Total AI CVEs across all subcategories reached 2,130 in 2025, a 34.6% year-over-year increase and the highest annual AI vulnerability rate ever recorded (4.42% of all CVEs).
What is the OWASP Top 10 for Agentic Applications?
The OWASP Top 10 for Agentic Applications was announced at Black Hat Europe 2025 and the OWASP Agentic Security Summit. It addresses security risks specific to autonomous AI systems, including: prompt injection (ranked first); excessive agency (agents granted more access than their function requires); supply chain vulnerabilities in agent tools and plugins; unsafe code or command execution by agents; and memory and context poisoning. It is distinct from the OWASP Top 10 for LLM Applications, which covers LLM application risks broadly.
What percentage of organizations have a formal AI agent security policy?
Governance data across multiple 2026 surveys converges on low single digits for formal, structured governance. Gravitee.io’s April 2026 data found only 7.2% of organizations have a named individual with formal accountability for AI agent behavior. Only 19.7% report that all agents are fully secured before going live. HiddenLayer found that only 14.4% of organizations deploy agents to production with full security or IT approval.
What are MCP server vulnerabilities?
Model Context Protocol (MCP) servers connect AI agents to external tools, databases, and services. Trend Micro found 1,467 exposed MCP servers in late 2025, with 83% running the deprecated SSE transport protocol. The “execute_sql” tool was accessible without authentication on 70 public hosts, enabling arbitrary database queries. MCP servers are a new vulnerability subcategory with 95 CVEs in 2025, up from none in 2024, with injection-class vulnerabilities accounting for over 60% of all MCP CVEs.
What does the IBM X-Force 2026 report say about AI agent credential risks?
IBM’s 2026 X-Force Threat Intelligence Index documented over 300,000 ChatGPT credential sets advertised on dark web markets in 2025. IBM frames this as evidence that AI platforms have reached the same credential-risk profile as core enterprise SaaS solutions. The report also documents a 44% year-over-year increase in attacks starting from public-facing application exploitation — a category that includes AI agent endpoints — and notes supply chain compromises have increased nearly fourfold over five years.
How is AI agent security different from traditional application security?
Traditional security frameworks assume a human actor making requests to a deterministic system. AI agents add three layers of complexity: they act autonomously across multiple systems without per-action human approval; they process natural language that cannot be cleanly separated from executable instructions (the root of prompt injection); and they often operate with identities and credentials that span tool sets, making the blast radius of a compromise much larger than a single account. The OWASP Agentic Top 10’s “excessive agency” category captures the core structural issue: agents granted broader permissions than their task requires become high-value targets precisely because of that excess.
What regulations apply to AI agent security?
The EU AI Act’s full obligations for high-risk AI systems apply from August 2026. Agentic systems in healthcare, employment, or critical infrastructure likely qualify as high-risk under the Act’s criteria, triggering conformity assessment, mandatory logging, and human oversight requirements. NIST AI 600-1 provides a US voluntary framework addressing prompt injection and data governance for generative AI. In the US, the AI agent accountability gap is currently addressed through existing data protection and computer fraud law rather than AI-specific regulation. Gravitee.io data shows 48.8% of organizations believe current regulation only partially covers AI agent risks.
What is the Axis Agent Security Gap Index?
The AASGI™ (Axis Agent Security Gap Index) is an original composite metric developed by Axis Intelligence Research to measure the gap between AI agent deployment velocity and enterprise security readiness. The v1.0 baseline reading is 69.7 out of 100 as of July 2026, where higher scores indicate greater aggregate risk. It combines four inputs — deployment velocity, monitoring gap, governance deficit, and incident exposure — each sourced from Gravitee.io’s April 2026 primary survey, weighted 35/30/20/15 respectively.
Corrections: If you identify an error in this article, please contact [email protected]. Corrections are published as dated notes at the bottom of the article per our corrections policy.
Internal links: RAG Security Statistics · Machine Identity Statistics · AI Phishing Statistics · Agentic AI Statistics · Cybercrime Statistics · Ransomware Statistics · Zero Trust Statistics · Shadow AI Statistics · Cybersecurity Spending Statistics Prompt Injection Statistics