Book a Call
Onnex
Book a Call
Your business has the answer.
Let Onnex find it.
LOADING
. . .
Completed
0

Verified in the first week of August 2026. Two frameworks moved days before publication, so re-check before any reprint.

SourceDateWhat it supports
IBM, Cost of a Data Breach Report 202629 July 2026Global average $4.99M, US average $11.50M, AI-enabled breaches ~$6M, prompt injection $5.89M, model inversion $6.07M, shadow AI $5.39M vs $4.63M, 43% shadow AI prevalence, 68% no completed AI governance policy, 247-day mean time to identify and contain, 1 in 4 malicious breaches AI-enabled. Base: 602 organizations breached March 2025 to February 2026.
Gartner, worldwide AI spending forecast19 May 2026$2.596 trillion total AI spending in 2026, up 47%. AI cybersecurity line item $51.347 billion, about 2%. Note the category bundles AI-for-security with securing-AI, so 2% is an upper bound.
MITRE ATLAS data repository, release 2026.0731 July 202616 tactics, 178 techniques (101 parent, 77 sub), 37 mitigations, 68 case studies. Monthly content releases.
OWASP GenAI Security Project, LLM Top 10 for 2026, v1.03 August 2026Prompt Injection first for a third year. Excessive Agency up to third. LLM08 is now Hidden Context Exposure. Rankings derived from analysis of around 10,000 real-world incidents.
OWASP Top 10 for Agentic Applications9 December 2025ASI01 to ASI10, used verbatim in the agentic article.
Hackett, Birch, Trawicki, Suri, Garraghan (Lancaster University and Mindgard), arXiv:2504.11168, LLMSec 20252025Up to 100% evasion against six production protection systems including Azure Prompt Shield and Meta Prompt Guard.
Schwinn et al. (TU Munich and Mila), arXiv:2603.06594February 2026LLM judge performance degrades to near random chance under red-teaming distribution shift, across 6,642 human-verified labels.
Maloyan and Namiot, arXiv:2601.17548January 2026Over 85% attack success against agentic coding assistants with adaptive strategies. Most defenses under 50% mitigation. Review of 78 studies, 42 techniques.
Li et al., RobustJudge, arXiv:2506.09443revised November 2025Up to 40% robustness swing from prompt-template choice. 15 attacks, 7 defenses, 12 models.
Meta, Llama Prompt Guard 2 model cards202519.3 ms for the 22M encoder, 92.4 ms for the 86M encoder.
NVIDIA developer blog, guardrail effectiveness and performance3 March 2025Three NeMo rails add roughly 530 ms end to end. First-party measurement of NVIDIA's own product.
TrueFoundry guardrail benchmark19 May 2026Azure PII 52.3 ms, OpenAI moderation 191.5 ms, Pangea prompt injection 358.7 ms. Published by a gateway vendor, not independent.
Sysdig, LLMjacking research2024 to 2025Roughly $50,000 of stolen compute in 4.5 days from one monitored proxy (observed). $46,000 per day is Sysdig's modelled ceiling, not a measured loss. Label it as modelled wherever it appears.
Anthropic, disrupting the first reported AI-orchestrated cyber espionage campaign13 November 2025GTG-1002. AI performed 80 to 90% of the campaign. ~30 targets. Humans intervened at 4 to 6 decision points. Succeeded in a small number of cases.
Anthropic, AI-enabled cyber threats mapped to MITRE ATT&CK3 June 2026832 banned accounts over 12 months. 67.3% used AI to write malware. Medium-risk-or-higher actors rose from 33% to 56%.
Five Eyes cyber agencies, "The AI shift in cyber risk"22 June 2026ACSC, CCCS, GCSB, NCSC, CISA, NSA. "The timeline is not years, it is months."
Varonis Threat Labs, SearchLeak, CVE-2026-4282415 June 2026One-click exfiltration from M365 Copilot Enterprise. Microsoft fixed 4 June 2026, rated critical.
Aim Security / Microsoft, EchoLeak, CVE-2025-32711June 2025First zero-click prompt injection in a production LLM system.
Zscaler ThreatLabz, indirect prompt injection targeting AI agents2 July 2026Two live campaigns. Four of 26 tested models executed fraudulent payments.
Sysdig, JADEPUFFER1 July 2026First complete LLM-driven ransomware. Self-corrected a failed login in 31 seconds. 1,342 Nacos config items encrypted. Key never transmitted.
Invariant Labs, MCP tool-poisoning disclosure1 April 2025Direct tool poisoning, rug pull and tool shadowing. Exfiltration succeeded even with user confirmation because arguments were hidden.
Microsoft, poisoned MCP tool descriptions warning30 June 2026MCPTox benchmark: poisoned-tool attacks effective at rates up to 72.8%.
NSA and CMU SEI, MCP security design considerations, CSI v1.0May 2026Eight named threat classes including output poisoning.
UK NCSC, "Prompt injection is not SQL injection"8 December 2025"It's very possible that prompt injection attacks may never be totally mitigated."
OpenAI, understanding prompt injections7 November 2025"A hard, open problem."
Anthropic, prompt injection defenses24 November 2025"No browser agent is immune to prompt injection."
Google DeepMind, advancing Gemini's security safeguards20 May 2025Baseline defenses "became much less effective against adaptive attacks."
European Commission, AI Omnibus enters into force27 July 2026Article 50 transparency applied 2 August 2026. Annex III high-risk deferred to 2 December 2027, Annex I to 2 August 2028. Article 99 penalties unchanged at €35M or 7%.
Colorado SB 26-189, signed by Governor Polis14 May 2026Repealed and replaced the Colorado AI Act. New ADMT transparency law effective 1 January 2027.
California SB 53, Transparency in Frontier AI Actin force 1 January 202610^26 FLOP threshold, $500M revenue for large developers, 15-day incident reporting, $1M per violation.
Executive Order, Eliminating State Law Obstruction of National AI Policy11 December 2025DOJ AI Litigation Task Force, FCC proceeding, FTC policy statement, BEAD conditions.
Acquisition timelineAug 2024 to Jun 2026Acquirer press releases plus Infosecurity Magazine M&A round-ups for Feb and Mar 2026. Only F5 / CalypsoAI disclosed a price ($180M). All other reported figures are press estimates and are omitted from the site.
AI-Sentinel internal benchmarks, v6.0June 2026Sub-20 ms overhead, average clean pass under 8 ms, rejections 2 to 4 ms, p99 2.2 ms at 200 rps, 2,000+ rps per instance, 1,237 automated tests, 160 mapped to MITRE IDs, 16 Rust crates, ~205,000 lines. Labelled as internal on every appearance.
MSP pilot resultsJune 202685.1% adversarial action rate across 1,000 payloads, zero confirmed exfiltrations across 3,000 live attacks, 100% of system-prompt extraction and cost-amplification probes blocked. Client unnamed at their request.

Ready to transform your business? Let’s talk.

Book a Call
© 2026 Onnex Global. All rights reserved.