IDS and IPS: where they sit, what they signal and where they stop
An IPS blocks, an IDS cannot; placement decides the view, and an alert is not a verdict.

An intrusion detection and prevention system (IDPS) watches traffic or host activity for signs of an attack; intrusion prevention also tries to stop it. The IDS/IPS split is not in detection — both analyse the same activity the same way — but in authority: an IDS alerts, an IPS may block, and blocking forces it into the path. So placement decides what the sensor can see and whether it can act; the alert is a rule match, not a verdict.
The mental model
observation signal response
┌───────────────────┐ ┌────────────────────┐ ┌────────────────┐
│ where the sensor │ match │ an alert: a rule │ │ passive: report│
│ watches traffic │ ───────▶ │ match, not a │ ─────▶ ├────────────────┤
│ inline or passive │ │ verdict │ │ inline: block │
└───────────────────┘ └────────────────────┘ └────────────────┘
what it can see what it means what it may do
Placement and meaning are separate: a sensor can match a signature exactly and still be wrong about what the activity means.
Terminology
- IDS / IPS / IDPS — a detection system observes and reports; a prevention system also acts; NIST uses IDPS.
- Sensor — the component that sees events; a network sensor sees traffic, a host agent one host.
- Inline — in the path; “the network traffic it is monitoring must pass through it”.
- Passive — “monitors a copy of the actual network traffic; no traffic actually passes through the sensor”.
- Alert / signature — the signal emitted when observed activity matches a pattern of known malicious activity.
- False positive / false negative — benign activity called malicious / malicious activity not called.
- Evasion — “modifying the format or timing of malicious activity so that its appearance changes but its effect is the same”.
Three detection methods, and what each cannot see
- Signature-based matches activity against patterns of known attacks: precise on what it covers, “largely ineffective at detecting previously unknown threats, threats disguised by the use of evasion techniques, and many variants of known threats”.
- Anomaly-based models normal behaviour and flags deviation: it “can be very effective at detecting previously unknown threats”, but benign deviation produces alerts, and the model can be walked: “an attacker can perform small amounts of malicious activity occasionally, then slowly increase the frequency and quantity of activity” until the attack is baseline.
- Stateful protocol analysis compares activity with a protocol model: “very resource-intensive”, and it “cannot detect attacks that do not violate the characteristics of generally acceptable protocol behavior”.
The two error types trade: “It is not possible to eliminate all false positives and negatives; in most cases, reducing the occurrences of one increases the occurrences of the other.”
Where the sensor sits decides what it sees
SP 800-94 gives two modes. Inline: traffic “must pass through it”; as NIST puts it, “The primary motivation for deploying IDPS sensors inline is to enable them to stop attacks by blocking network traffic”. Passive: the sensor sees a copy, so passive techniques “typically provide no reliable way for a sensor to stop the traffic from reaching its destination”. The rule: “Generally, organizations should deploy sensors inline if prevention methods will be used and passive if they will not.”
Sensors belong at “the divisions between networks” — where a firewall would sit; an inline one is often on the more secure side, a passive one watches segments “such as activity on a demilitarized zone (DMZ) subnet”. The divisions’ policy, the DMZ design and the mechanics of a copy are the segmentation, DMZ and packet-capture sheets.
Three things decide what it sees: encryption — it “cannot detect attacks within encrypted network traffic”; load — it “may be unable to perform full analysis under high loads”, and inline “dropping packets under high loads causes disruptions in network availability”; and direction — perimeter sensors watch north–south while “over 75 % of network traffic is now east-west” and “Any threat that is already inside of a network can move laterally and remain undetected for days or even months.”
An alert is a signal, not a verdict
An alert claims observed activity, not intent. The trade-off is a cost: “Most IDPSs require at least some tuning and customization to improve their detection accuracy, usability, and effectiveness”. An alert is also weak evidence alone: where malicious activity is rare, even a good sensor returns mostly false positives, because benign traffic is far more abundant; an alert becomes informative only with context supplied by the sibling sheets: event logs and severity (syslog), flow records without payload (NetFlow/IPFIX), packets where a capture exists. Alert volume is an attack surface: a burst buries the real event.
Where it stops
- It cannot see what placement hides. Encrypted traffic, traffic it never receives, and activity with no signature produce no alert; no alert is not evidence of absence.
- It can be evaded. Organisations should “use network-based IDPSs that can compensate for the use of common evasion techniques”; a sensor that cannot normalise an encoding the target understands will miss it.
- An inline blocker is an availability risk. It sits in the path: high load, a corrupt update or one bad rule become outages.
- It must be managed. Sensors are best run on a management network; one without addresses on its monitoring interfaces is “operating in stealth mode”; and a fresh sensor is “likely to generate a large number of false positives”.
- No single sensor is enough. “In many environments, a robust IDPS solution cannot be achieved without using multiple types of IDPS technologies.”
Level and prerequisites
L4 — advanced. It assumes the L3 sheets on segmentation and trust domains, the DMZ and service networks, and the logging sources, none of which it re-explains.
Where to go next
References
- NIST — Guide to Intrusion Detection and Prevention Systems (IDPS) (SP 800-94, Karen Scarfone and Peter Mell, February 2007) — the definition of intrusion prevention (line 295); false positives, false negatives, tuning and evasion (lines 640–655); the three detection methodologies and their limits (lines 667–810); inline and passive sensor modes and placement (lines 1689–1778); the phased-deployment guidance (lines 1185–1187); the management and stealth-mode guidance (lines 1228, 2158–2159); the deployment-time limits: encryption, high load and alert volume (lines 2010–2050); the need for multiple technologies (lines 3799–3800).
- NIST — Guide to a Secure Enterprise Network Landscape (SP 800-215, November 2022) — the north–south focus of perimeter IDS/IPS and firewalls, the east-west share of enterprise traffic, and lateral movement inside a network (lines 500–508).
- RFC 4949 — Internet Security Glossary, Version 2 (R. Shirey, August 2007) — the definitions of intrusion detection (lines 9254–9256), intrusion detection system (lines 9269–9272), anomaly detection (lines 982–984) and misuse detection (lines 10774–10777).