AI security

AI threat tracker

How attacks on AI software, new vulnerabilities in AI packages, reported AI incidents and threat actors’ use of AI change month by month, from public sources.

Updated 21 Nov 2026

AI software under active attack

CISA added 10 vulnerabilities in AI software to its list of exploited vulnerabilities in the last 12 months, up from 1 in the 12 months before.

Show the numbers
Monthadded
Dec 20240
Jan 20250
Feb 20250
Mar 20250
Apr 20250
May 20251
Jun 20250
Jul 20250
Aug 20250
Sept 20250
Oct 20250
Nov 20250
Dec 20250
Jan 20260
Feb 20260
Mar 20261
Apr 20260
May 20262
Jun 20261
Jul 20262
Aug 20262
Sept 20261
Oct 20260
Nov 20261

By kind of software

Kind of softwareLast 12 months12 months before
Agent builders and frameworks51
AI apps, gateways and interfaces40
ML operations and data tools10

Latest additions

CVE-2026-40217BerriAI LiteLLM Unprotected Alternate Channel Vulnerability18 Nov 2026
CVE-2026-59822BerriAI LiteLLM Improper Authentication Vulnerability2 Sept 2026
CVE-2026-64849MLflow Server-Side Request Forgery Vulnerability19 Aug 2026
CVE-2026-9198IBM Langflow Code Injection Vulnerability4 Aug 2026
CVE-2026-0770Langflow Inclusion of Functionality from Untrusted Control Sphere Vulnerability21 Jul 2026
CVE-2026-55255Langflow Authorization Bypass Through User-Controlled Key Vulnerability7 Jul 2026
CVE-2026-42271BerriAI LiteLLM Command Injection Vulnerability8 Jun 2026
CVE-2025-34291Langflow Origin Validation Error Vulnerability21 May 2026

What this does not show

The list only includes vulnerabilities CISA has reliable evidence are being exploited, and it leans towards software used by US government agencies.

Hosted AI services such as chatbots rarely get CVE numbers, so attacks on them do not appear here.

An entry is counted when the product is on ResearchNode’s list of AI software. Entries that only mention AI in their description are checked by hand first.

New vulnerabilities in AI packages

551 vulnerabilities were published for the 62 AI packages on the watchlist in the last 12 months, up from 222 the year before. 332 are rated critical or high, and 14 have no rating yet.

new vulnerabilitiescritical or high

Show the numbers
Monthnew vulnerabilitiescritical or high
Dec 202411
Jan 202532
Feb 202531
Mar 20258359
Apr 2025168
May 2025168
Jun 202554
Jul 20252111
Aug 202586
Sept 20253021
Oct 20252213
Nov 20251210
Dec 20252524
Jan 20261515
Feb 20262315
Mar 20264536
Apr 20265838
May 202611073
Jun 20266229
Jul 20264422
Aug 20265335
Sept 20264914
Oct 20266428
Nov 202611

By kind of software

Kind of softwareLast 12 months12 months beforeMost often affected
AI apps, gateways and interfacesopen-webui · 13917739open-webui · 139
Agent builders and frameworksflowise · 8114049flowise · 81
Model servers and inferencevllm · 8211850vllm · 82
ML frameworks and librariestransformers · 134347transformers · 13
ML operations and data toolsmlflow · 323926mlflow · 32
MCP servers and SDKsmcp · 92510mcp · 9
Vector databaseschromadb · 491chromadb · 4

Latest critical vulnerabilities

CVE-2026-116692MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modules27 Oct 2026
CVE-2026-115915vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE26 Oct 2026
CVE-2026-73487Flowise Prompt Injection to RCE and SSRF via CSV/Airtable Agent Python Validator Bypass7 Oct 2026
CVE-2026-119978Flowise NodeVM sandbox escape via puppeteer allowlist - authenticated RCE and arbitrary file read via Chromium7 Oct 2026
CVE-2026-105697Langflow: OS command injection (RCE) via arbitrary command in MCP stdio server configuration7 Oct 2026
CVE-2026-10561Langflow: PythonREPLComponent executes unsandboxed Python code, enabling authenticated RCE and privilege escalation6 Oct 2026
CVE-2026-9205Langflow: Weak Fernet Key via random.seed()5 Oct 2026
CVE-2026-8505Langflow: Unauthenticated Flow Execution via Webhook Authentication Bypass5 Oct 2026

What this does not show

More vulnerabilities can mean more people are looking, not that the software got worse. The exploited list above is the better sign of real attacks.

Only packages published on PyPI, npm and Go are counted. AI apps installed from source, and hosted services, are missing.

The same flaw listed by GitHub, PyPI and NVD is counted once, by its CVE number. Severity is GitHub’s reviewed rating where there is one.

Reported AI incidents

The AI Incident Database lists 373 incidents that happened in the last 12 months, down from 446 in the 12 months before. 72 of them are classified as malicious actors and misuse, against 234 the year before.

Source: Responsible AI Collaborative’s AI Incident Database export from 16 Nov 2026. ResearchNode groups the records by incident date and MIT risk category. Incident data: CC BY-SA 4.0.

The dashed end and open nodes are the last three months, which are still being filled in. Incidents are often added weeks after they happen.

incidentsmalicious actors and misuse

Show the numbers
Monthincidentsmalicious actors and misuse
Dec 20242618
Jan 20253319
Feb 20254730
Mar 20254626
Apr 20254827
May 20253418
Jun 20254525
Jul 20253915
Aug 20253516
Sept 2025229
Oct 20254219
Nov 20253314
Dec 20254519
Jan 20263917
Feb 2026339
Mar 20262911
Apr 2026297
May 2026425
Jun 2026320
Jul 2026390
Aug 2026170
Sept 2026300
Oct 2026250
Nov 202640

By risk area, last 24 months

Risk area (MIT AI Risk Repository)IncidentsShare
Malicious actors and misuse30637%
Not yet classified27934%
Misinformation8510%
AI system safety, failures and limitations547%
Privacy and security364%
Discrimination and toxicity344%
Human-computer interaction192%
Socioeconomic and environmental61%

Most recent incidents

Incident 1827Autonomous Haul Truck Reportedly Went Down an Embankment After Communications Loss at Copper Mine in Antofagasta, Chile12 Nov 2026
Incident 1825Autonomous a2z ROii Bus Struck Roadside Stone Wall During Passenger Trial in Toyooka, Japan9 Nov 2026
Incident 1809Scotland's Education Agency Published Purportedly AI-Edited School Curriculum Papers with Apparently Nonexistent References5 Nov 2026
Incident 1815Attacker Reportedly Used BOSSAI AI Agents to Extract Emergency Dispatch Records in Taipei, Taiwan4 Nov 2026
Incident 1795Attacker Reportedly Used AI Agents to Breach Blockchain Analytics Firm Chainalysis and Access Customer Data30 Oct 2026
Incident 1817Hackers Reportedly Used Internal AI Tool in $6.34 Billion Breach of Cryptocurrency Exchange Binance29 Oct 2026

What this does not show

Reported is not the same as happened. The count follows what news media and submitters notice, and media attention to AI keeps growing.

The database is maintained by volunteer editors, so its coverage depends on what they review.

Risk areas come from the database’s own MIT classification, which is partly done by a language model.

How threat actors use AI

No records yet. Records appear here once a report has been read and checked.

What this does not show

Each company only sees misuse of its own AI products, and publishes on its own schedule. A jump usually means a report came out, so this is shown as records, not a trend line.

Names, countries and attributions are the publishing company’s.

Sources

CISA Known Exploited VulnerabilitiesThe official list, read twice a day. The GitHub copy CISA publishes is used if the main address fails.21 Nov, 02:55
OSVAdvisories for the AI packages on the watchlist, from GitHub, PyPI and Go, read once a day.21 Nov, 02:54
AI Incident DatabasePublished weekly incident export. ResearchNode checks daily for a new snapshot.20 Nov, 08:14
Company threat reportsPublished reports from AI companies about misuse of their products, read with the help of a local AI model and checked by hand.21 Nov, 03:33

The tracker updates automatically from these sources. Spotted something wrong? See Corrections.