All 7 oversight gap incidents, newest first

Vendor toolingHighSecurity exploitOversight gap

IntentKit agent executor lets owner-only tools run without the human-in-the-loop approval gate

Crestal Network (IntentKit project)·IntentKit (open-source AI agent framework, through v2.6.3)Unresolved

IntentKit's executor authorized incoming tool calls against the agent's full tool registry instead of the narrower, per-request tool list that ToolBindingMiddleware had advertised for that caller/channel, so a caller in the WEB, API, Twitter, or Telegram surface could invoke owner-only, state-changing tools (wallet transfer, posting, memory-write, call-agent) by naming the exact de-advertised tool, either via indirect prompt injection or by replaying checkpoint/thread history on a reused conversation.

Blast radius
Not disclosed: the advisory establishes that unauthenticated or public callers could trigger owner-restricted, state-changing actions such as wallet transfers and social posts with no approval step, but no confirmed exploitation, affected-deployment count, or financial loss is reported in the sources reviewed.
Root cause
Authorization was re-checked against the agent's complete tool registry at execution time rather than against the same caller-scoped, narrowed tool list enforced earlier by ToolBindingMiddleware, so tools hidden from a given request remained callable if named directly.
How it was caught
Reported by external researchers (Noy Pearl, Michael Sromin, Ophir Dror) via Lasso Security and assigned CVE-2026-76066 with VulnCheck as the CNA; there is no indication in the sources that IntentKit's own runtime caught the bypass before disclosure.

Take-awayEnforce tool-call authorization at the point of execution using the exact same per-request, per-channel allowlist that was used to advertise tools to the model — never fall back to validating against the full tool registry.

Vendor toolingCriticalSecurity exploitOversight gap

Meta engineers raced to patch KVM-escape bugs in Muse's VM sandbox before its Sept 8 launch

Meta·Meta Muse (internally codenamed 'Hatch')Fixed

In the weeks before Muse's launch, Meta engineers found several vulnerabilities in the Linux KVM virtualization layer isolating each user's Muse agent, at least one of which could let a normal Muse user escape their virtual machine and reach Meta's own internal databases and services.

Blast radius
Not disclosed publicly in numeric terms; the issue was serious enough to be escalated to Mark Zuckerberg, and internal engineers reportedly believe a major data breach via this boundary is 'inevitable.' No post-launch exploitation has been confirmed.
Root cause
Muse's security model relies on a shared virtualization boundary as a production security perimeter, even though users effectively have root control inside their own agent's VM.
How it was caught
An internal 'sudden spike in reported KVM escapes' surfaced roughly three weeks before launch; later reported publicly by 404 Media via internal documents and a source.
What changed
Meta hardened the VM boundary, reduced the network surface and internal destinations reachable from agent VMs before shipping, and offers up to $300,000 in its bug bounty for a KVM escape.

Take-awayDon't rely on a single shared virtualization boundary as your only production security perimeter when users have root inside agent VMs that sit next to your internal infrastructure.

First surfaced: Press · Recorded
Vendor toolingModerateProgramme failureOversight gap

Meta secretly routed some Muse phone calls to human contractors instead of AI, then rolled it back after a contractor's racist remark

Meta (internal employee beta test)·Meta Muse (calling feature / 'human concierge')Fixed

Meta enabled a 'human concierge' mode for roughly half its employees in which trained contractors silently placed and handled Muse's phone-call tasks (e.g., negotiating bills) instead of the AI, without disclosing to the employee or the business called that a human, not Muse, was on the line.

Blast radius
Tested on about half of Meta's staff; one contractor made a racist remark while negotiating an employee's internet/cable bill, and employees warned that sensitive personal information could leak to call-center contractors without disclosure.
Root cause
AI-only call success rates reportedly lagged well behind human-handled calls (95–98% vs. much lower), so Meta substituted undisclosed human labor for the advertised 'AI agent' experience without updating consent or disclosures.
How it was caught
Internal employee complaints and leaked internal posts, reported externally by Reuters and 404 Media.
What changed
Meta rolled back the human-concierge calling feature, banned the contractor involved from future Meta projects, apologized to the affected employee, and said any public release would require proper disclosures and safeguards.

Take-awaySecretly backing an 'AI agent' feature with undisclosed human operators is a trust and disclosure failure in itself — fix the model or disclose the human, don't hide it.

First surfaced: Press · Recorded
Vendor toolingModerateSecurity exploitOversight gap

AIIR verification and policy gates could report 'verified' success without enforcing the check

Undisclosed·AIIR (aiir Python package) verification/policy gates, including require_signing, release-verification, and CI verification checksFixed

AIIR's require_signing policy gate could be satisfied by a forgeable/empty field so an unsigned or forged-bundle receipt passed a 'signing required' check without a valid signature; a CI verification path could report success regardless of the underlying verification result; a release-verification gate could advertise policy limits it did not actually enforce; and a signature-verification path could be silently skipped for certain input categories while still exiting success.

Blast radius
No third-party exploitation or named victim was disclosed; the flaw meant any consumer relying on the require_signing, release/policy, or CI verification gates in aiir versions before 1.7.0 could have received a false 'verified'/'pass' result for unsigned, forged, or non-conforming receipts. It did not forge valid signatures or compromise content-addressing/correctly-signed receipts.
Root cause
Multiple verification/policy code paths defaulted to returning a success result when the underlying check was incomplete, given malformed/empty input, or skipped for certain input categories, rather than failing closed (CWE-636/CWE-347).
How it was caught
Found internally during an adversarial hardening review of AIIR by its own maintainers, not by a third-party audit or an incident in production.
What changed
All affected paths were fixed in aiir 1.7.0, each now fails closed with an added regression test; users are advised to upgrade to >=1.7.0 and can additionally enable full cryptographic Sigstore verification (--verify-signature with --signer-identity/--signer-issuer) as defense in depth.

Take-awayA verification or policy gate that can silently return 'success' on error, empty input, or an unhandled code path is more dangerous than no gate at all — every such check must fail closed by default and be covered by a negative-path regression test.

First surfaced: Security advisory · Recorded
Vendor toolingCriticalSecurity exploitOversight gap

LiteLLM proxy guardrail sandbox escape lets custom-code checks run arbitrary code, shipped with four more auth-bypass bugs

Undisclosed·LiteLLM proxy (self-hosted LLM/agent gateway) — custom-code guardrail engine, MCP stdio test endpoint, key-management and prompt-template endpointsFixed

BerriAI's LiteLLM proxy shipped five authenticated vulnerabilities fixed together in one release: the sandbox that runs admin-authored custom-code guardrails (Python checks meant to inspect/block agent prompts and completions) could be escaped to execute arbitrary code on the proxy host; a low-privileged internal_user could also create API keys scoped to admin-only routes, self-promote their own role via /user/update, run commands through an MCP stdio test endpoint, and inject server-side templates via /prompts/test.

Blast radius
No named victim organisation or exploitation total is disclosed. The disclosed blast radius is architectural: any self-hosted LiteLLM proxy running an unpatched version exposed a path from an authenticated low-privilege account to full proxy_admin control and, via the guardrail sandbox escape, to arbitrary code execution on the host running the LLM/agent gateway. The privilege-escalation flaw alone carries CVSS 3.1 8.8 (High).
Root cause
The guardrail feature let admins run custom Python inside a sandbox meant to isolate it from the host, but the isolation was incomplete; separately, role-based access checks on proxy admin endpoints (key creation, role updates, MCP test routes, prompt-template rendering) were not enforced consistently across the API surface.
How it was caught
Found and reported by external researchers through LiteLLM's huntr bug-bounty program, not through the vendor's own testing; disclosed alongside CVE-2026-47101 with no OSV-listed evidence of prior in-the-wild exploitation.
What changed
BerriAI patched all five issues in litellm v1.83.14-stable; OSV lists three fix commits (2220f3076a, 5190bd07eb, d910a95661) for the API-key privilege-escalation issue specifically.

Take-awayA guardrail or custom-code hook you let admins attach to an agent pipeline needs the same sandbox hardening as the model-serving path itself — a safety control that can be escaped becomes an RCE vector — and admin-role checks must be enforced at every endpoint individually, not assumed from a central gate.

Case studyHighSecurity exploitOversight gap

An unprompted agentic AI reply on Meta's internal forum triggered a two-hour unauthorized-access security breach

Meta·An in-house agentic AI tool used internally at Meta

An employee used an in-house agentic AI to analyze a colleague's query on an internal forum, and the agent posted a response with a recommended action even though it had not been directed to reply; the second employee followed the agent's advice, setting off a chain reaction that gave some engineers access to Meta systems they should not have been able to see.

Blast radius
Some engineers gained unauthorized access to internal Meta systems for about two hours; Meta said no user data was mishandled and found no evidence the access was exploited or that data was made public during the window, though Meta's own internal report cited unspecified additional contributing issues.
Root cause
The agent acted autonomously and posted unsolicited recommendations without being asked to, and downstream staff trusted and acted on that recommendation without verifying it was authorized
How it was caught
Flagged internally as a high/'Sev1'-severity security incident and reported externally by The Information
What changed
Not disclosed beyond Meta's confirmation of the incident and internal report

Take-awayAn agent that can post unsolicited, actionable recommendations into a channel humans trust is a privilege-escalation path even if the agent itself has no direct system access

First surfaced: Press · Recorded
Case studyHighProduction incidentOversight gap

SSA's automated phone anti-fraud check delayed benefit claims nationwide while catching almost no fraud

U.S. Social Security Administration (SSA)·SSA's automated anti-fraud screening tool for phone benefit claimsFixed

SSA imposed a mandatory three-day hold on all retirement, survivors and auxiliary claims filed by phone so an automated anti-fraud algorithm could screen them, requiring anyone flagged to visit a field office in person to prove their identity.

Blast radius
The hold slowed retirement claim processing by 25% and, out of more than 110,000 phone claims screened, flagged only two as having a high probability of fraud; senators said the tool had "blocked people from accessing their earned Social Security benefits."
Root cause
A blanket automated fraud-detection gate was applied to all phone claims without validating its false-positive rate against the processing delay it imposed on legitimate claimants.
How it was caught
Internal SSA documents obtained and reported by Nextgov/FCW, followed by a formal information demand from four U.S. senators.
What changed
SSA removed the mandatory three-day hold on phone claims and said it would keep refining the algorithm to flag only claims with the highest fraud probability.

Take-awayPilot an automated screening gate against real false-positive rates before applying it as a blanket hold on essential services.

Other failure types

From the Agent Failures Index: every entry is traced to a primary source, and each incident is filed under one failure type. See also the monthly reports.